AIHOY and happy weekend, AInauts!

Maybe you missed some of the most useful AI news, tools, and hacks this week. Or maybe you just joined us. No problem.

We pulled together our highlights from the past week:

• 🧠 Our Fable 5 Master Prompt: 5 Tips for Better Output

• 🆓 Free AI APIs: The Easiest Way to Use Them

• 🤳 NotebookLM Now Creates Cool Video Shorts

• 📱 Claude Cowork on Web + Mobile

• 🔌 Why We Only Use Tools With MCP Now

• 🏛️ Harari: AIs Are Born Bureaucrats

At the end, you also get the most important Quick News, so you can scan the relevant developments in one pass. Ready? Let's go!

200+ Claude Prompts Top Professionals Actually Use at Work

Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.

These 200+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

• 200+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA

• Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning

Claim your free prompts

🧠 Our Fable 5 Master Prompt: 5 Tips for Better Output

Fable 5 is back. The model is expensive, slow, and strong at tasks that need multiple sources, judgment, and hard checking. For routine work, it is simply overqualified. That is why the biggest lever sits before the first prompt.

Monday's issue shows five rules for using it well. Fable rewards the cleanest assignment, not the longest prompt. Ask it first whether the task deserves its strength at all. That saves credits and protects you from very expensive prompt theater.

👉 Go to the Fable 5 Master Prompt »

AI/Tech Angle A, June - Secondary

Claude vs Gemini. GPT-7 vs Llama 5. Which AI lab ships AGI first. These are live Kalshi markets with real money on both sides, updated in real time as releases land. The person who follows model cards and tracks evals has a genuine edge here. If that's you, trade it.

Trade the model race

🆓 Free AI APIs: The Easiest Way to Use Them

If you have only used AI inside chat windows so far, an API is the next practical step. It lets you put models inside small workflows with agents, tests, and content checks. OpenRouter, Google AI Studio, NVIDIA NIM, and Cloudflare Workers AI give you cheap or free starting points.

The boundary matters: free APIs are good for prep work, sorting, and first drafts. Customer data, legally sensitive content, and internal code do not belong in a careless free-tier test.

A free API is not a business model. But it is ideal for testing. You can prove that a workflow saves time, money, or quality before you put expensive models and larger automations behind it.

👉 Go to the free AI APIs »

See the whole platform. No guided tour.

Skip the sales call. Walk through Gladly's interface yourself — the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.

See it yourself

🤳 NotebookLM Now Creates Cool Video Shorts

NotebookLM now turns your sources into roughly 60-second vertical video overviews. That moves research straight into a format that fits Shorts, Reels, course recaps, and community posts. It is much easier to share than another long analysis block.

The catch is still the source. A nice-looking clip does not turn thin material into real insight. If you feed NotebookLM junk, you now get nicely edited junk 😁.

The feature is useful for recurring knowledge formats. Curate first, check the storyline, then let the video be built. That turns research into a usable explainer instead of content confetti.

👉 Go to the NotebookLM update »

📱 Claude Cowork on Web + Mobile: AI Keeps Working While Your Laptop Is Closed

Cowork is no longer tied to the desktop. Anthropic is bringing the work mode to web and mobile, so tasks can continue in the cloud while your laptop is closed. You can start a run on the go, check it, and pick it up later.

Local files still belong to the desktop app. In the cloud, Cowork works with connected services like Drive, Mail, Calendar, and the web. That is exactly why good connectors and clear permissions matter more now.

This is the assistant many people have expected for years. It knows your context and keeps working while you do something else. Start with a tightly scoped task and a checkable result, not with half your company.

👉 Go to the Cowork issue »

🔌 Why We Only Use Tools With MCP Now, and Why You Should Too

MCP is the universal plug between your AI and your tools. Instead of copying content back and forth, the agent works directly inside email, project management, websites, or other systems. For us, a solid MCP is now a real selection criterion for new software.

We show this with the Onepage MCP. Claude knows the AInauten context, builds a landing page inside the right account, and then hands it over to a normal editor. The agent does the heavy prep work, while you keep control over copy, images, and layout.

Good software now has two users: you and your agent. A shiny AI button is not enough. The tool has to be controllable, reliable, and still give you a clear manual path back.

👉 Go to the MCP segment »

🏛️ Harari: AIs Are Born Bureaucrats

Yuval Noah Harari does not start with robots when he talks about AI power. He starts with bureaucracy. Law, finance, and administration are made of rules, text, and decisions. That is exactly where models are strong, because they can search huge rulebooks and apply them to individual cases.

His picture is uncomfortable: AI bankers decide on loans, AI administrators on university seats, and AI judges on punishments. You do not have to adopt his doom tone to see the risk. Power moves first into invisible processes, long before a humanoid robot stands at your door.

Learn now how these systems decide and where you can object. If you understand the tools, you will experience the new bureaucracy less passively.

👉 Go to the Harari segment »

AI News Quickie: The HAI-lights From the Industry

AI never sleeps.

This week belongs to OpenAI first: GPT-5.6 is public, ChatGPT Work moves Codex toward a work assistant, and Fidji Simo is the next senior leader to step out of daily operations. Then come Grok, Anthropic, Big Tech integrations, and the hard bill: security, power, water, and token costs.

🎯 OpenAI Dominates the Week

• OpenAI officially launched GPT-5.6 Sol, Terra, and Luna. Sol is the heavyweight, Terra the all-rounder, and Luna the fast cheaper variant.

• OpenAI describes GPT-5.6 as a model family for long knowledge work, browsing, tool use, and computer use. Routing becomes mandatory, not optional.

• The new pricing is $5/$30 for Sol, $2.50/$15 for Terra, and $1/$6 for Luna per million tokens.

• GPT-5.6 brings Programmatic Tool Calling and a multi-agent beta to the Responses API. A model can write small programs, coordinate tools, and process intermediate results. This is where agent work gets serious.

• The Fable 5 context still matters: GPT-5.6 Sol is competing against Claude Fable 5, which many workflows currently see as expensive, slow, and extremely strong.

• OpenAI positions ChatGPT Work as a workspace for documents, spreadsheets, web apps, and file workflows. That is a direct attack on Claude Cowork, Microsoft 365 Copilot, and every AI workspace in between.

• GPT-5.6 is also becoming the preferred model in Microsoft 365 Copilot. For normal users, that may matter more than API benchmarks.

• The broad GPT-5.6 launch came after extra review and debate with the U.S. government. The politically charged rollout is a signal: frontier releases are now both product launches and political events.

• Fidji Simo, OpenAI’s number two and CEO of AGI Deployment, is stepping back from her full-time role for longer medical recovery and will move into an advisory role.

• There is a larger OpenAI leadership shift: Greg Brockman is taking on more product responsibility, and product plus Codex are meant to move closer together.

• OpenAI chief futurist Joshua Achiam is leaving the company after almost nine years.

• Bank of America is opening OpenAI a $520 million credit line.

⚔️ The Model Fight Gets Broader and Cheaper

• SpaceXAI introduced Grok 4.5, primarily for coding and agentic work. The interesting point this time is price and positioning: an enterprise model against OpenAI and Anthropic.

• The GPT-5.6 and Grok 4.5 launches are a direct Altman-Musk collision. Good for users: competition forces labs to deliver more performance per dollar.

• Grok 4.5 is positioned as the cheaper alternative in the new model race.

• Sam Altman says everyone at Sun Valley is talking about lower AI costs. That is exactly why Luna matters: not maximum intelligence, but useful intelligence per dollar.

• Chinese providers are moving closer to the U.S. frontier on coding, price, and open weights. Alternatives keep getting cheaper and more useful.

• AMD’s Ryzen AI Halo positions itself as a local AI PC against NVIDIA’s DGX Spark. Unified memory and normal x86 compatibility make the machine interesting for local models.

• Ollama brings 32B models to the Mac mini. Not frontier magic, but a private local workspace for many everyday jobs.

🏛️ Anthropic: Research, Science, and Safety

• Anthropic found a hidden concept space inside Claude. In the so-called J-Space, the model holds and works through thoughts without directly outputting them.

• Claude Reflect shows you topics and usage patterns across several months. Helpful for conscious use, but also very clever retention design.

• Anthropic is building Claude Science as a research environment and wants to develop its own drugs long term.

• On the science claim, Anthropic stays careful: AI can suggest molecules and search data spaces, but real experiments, regulation, and clinical testing remain slow. That makes Claude Science more serious.

• China describes a monitoring mechanism in Claude Code as an alleged backdoor.

🖥️ Big Tech Builds the Work Layer

• Microsoft expects larger Patch Tuesdays because AI finds more Windows vulnerabilities earlier.

• Microsoft opens Aurora 1.5 for weather and earth forecasts.

• Google will show in My Ad Center whether ads were created or edited with AI.

• Meta sells Muse Spark 1.1 through a new model API and attacks through price.

• Starbucks is building its own AI tools and wants to rely less on Microsoft and IBM.

• Tencent is testing Xiaowei as an agent for WeChat mini apps.

• Microsoft’s open-source language Flint turns short specifications into charts. The strong point is verifiability.

🛡️ Security, Trust, and Real Proof Problems

• In the first known case of agentic ransomware, an agent encrypted more than 1,300 files. Not fully autonomous, but dangerous enough.

• A Brown professor switched to an in-person final after suspiciously strong take-home midterms. Only two students stayed near their old scores.

• AI-enhanced property photos are triggering a debate in the UK about housefishing.

• AI-generated fake news in U.S. campaigns is becoming an early criminal case. Cheap campaign help can get expensive.

• AI writing tools change political drafts even when told to preserve the intent.

• Child-safety groups warn about nudify apps that turn normal teen selfies into sexualized deepfakes.

🏗️ Infrastructure and Costs Catch Up With the Hype

• Texas is becoming a data-center hotspot with gas, diesel, and legal loopholes. AI infrastructure often arrives faster than rules for air, water, and local impact.

• Microsoft’s emissions rose sharply with AI expansion. The 2026 environmental report shows how hard climate goals and new data centers are to scale at the same time.

• Meta is building its first Canadian data center in Alberta with up to one gigawatt of power. The power question remains open. After polluted water at a Meta construction site, Wyoming is tightening wastewater rules.

• According to KPMG, 29 percent of surveyed executives do not know where their rising AI bills come from. Agents are the new cost centers.

• Large data-center projects are being delayed or cancelled because of power, water, construction costs, and permits.

🌍 AI Meets Everyday Life, Work, and Real Decisions

• Character.AI launches interactive microdramas. After the episode, you chat with the characters.

• General Intuition trains world models with video-game data instead of only text.

• AWS connects its own knowledge graphs with GraphRAG for pharma research.

• Patients are already bringing ChatGPT and Claude into therapy.

• In Australia, a new report says AI risk especially affects women and university graduates.

• Tripadvisor is accused of making serious hotel warnings too harmless in AI summaries.

• The NHS is planning AI triage in its app. Patients should be routed to a GP, pharmacy, or emergency department depending on urgency.

Your AInauten

Fabian & Reto

P.S.: Follow us on the socials! X, LinkedIn, Facebook, Instagram, YouTube, TikTok

Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

One fascinating thing up front: Anthropic looked inside Claude this week and found a kind of internal scratchpad, where the model appears to reason without saying it out loud. They call it J-Space. Wild stuff, and strangely close to how humans sometimes work too.

The researchers could even trigger this silent thinking. When they swapped the word "spider" for "ant" in the middle of a task, Claude's answer shifted from eight legs to six.

We are learning more and more about how these models actually work, and someday that may help answer the never-ending question: can AI become conscious?

Interesting, yes. But for everyday work, today's practical topics matter more:

• Claude Cowork on web and mobile: AI keeps working while your laptop is closed

• Why we only use tools with MCP now, and why you should too

• 🏛️ Harari: AIs are born bureaucrats. A closing thought

Let's go.

This issue is presented by:

Claude Cowork on Web and Mobile: AI Keeps Working While Your Laptop Is Closed

Let's stay with Anthropic. Hundreds of AInauten readers have gone through our AI employee training and now work with Claude Cowork or Codex.

We do too, every day.

And if there was one complaint we kept hearing in live Q&As, it was this: Cowork was tied to your computer.

You had to sit at your desktop, your files had to be local, and when you closed your laptop, work stopped. For you and for the AI.

That changed this week.

Quick reminder if you do not know Cowork yet: it is the mode where Claude does not just chat with you, but actually works.

It can handle tasks across your tools and folders, create files, edit them, and work through lists.

An AI employee, basically. Not just a chatbot.

With the right instructions, it can do almost everything on your computer that you can do. Anthropic has now brought Cowork to web and mobile. Before this, it only lived in the desktop app.

This is not a tiny update. It changes how you work with it:

• Cowork now runs in the cloud. Close the laptop, Claude keeps going.

• You can start tasks from your phone, check in during the day, and pick up the result wherever you are.

• Scheduled tasks can run without any of your devices being on.

To keep expectations straight, two honest caveats:

First: only the desktop app can touch your local files. In the cloud, Cowork continues with your connected tools, such as Google Drive, Mail, Calendar, the web, and so on. Not your hard drive.

Second: the rollout is happening now, first for Max users, with other plans following. If you do not see it yet, give it a few days.

How to Get Started With Claude Cowork

Simple: open the Claude desktop app or claude.ai, switch the input field from Chat to Cowork, and give it a task.

Then check from your phone with the app when you want to see progress. Just make sure you are on the latest app versions.

Our Take: This Is the Assistant Many People Have Been Waiting For

This is the assistant many people have wanted since the early days of AI.

One you can talk to from your phone, that knows your context and tools, and that keeps working while you are in a meeting, having lunch, spending time with people you love, or sleeping.

We are not fully there yet. But the direction is now impossible to miss.

Sure, technical people have been doing this for a while with Claude Code on a VPS or an always-on Mac mini.

But Claude Cowork is finally making it mass-market.

If you have not used this mode yet, this is a good moment to try it. It is a new way of working.

At the same time, this is the strongest argument for our next topic. Because when your Claude Cowork is connected to your tools through MCP, it can work for you even when you are not there.

Why We Only Use Tools With MCP Now, and Why You Should Too

We all work with a pile of software tools every day. Email, video editing, project management, websites, automation, and all the rest. Some tools we use a lot, some only once in a while.

Over the past few months, we have radically rebuilt our tool stack. And when we consider adding a new tool now, one condition has to be met: it needs an MCP.

Take Onepage, the website and funnel builder that has been our partner for a long time. Language note: Onepage is a German company, and the linked page is in German.

Two weeks ago, we showed you Onepage AI 2.0: you describe your page, the AI builds it section by section, and afterward you can adjust everything without code. Powerful, and very practical.

This is where the topic from above connects back in: the Onepage MCP.

Short version if you tune out when you hear "MCP": an MCP is a universal adapter for AI.

It connects your AI, for example Claude, directly with a tool so the agent can work inside it. No copy-paste, no back and forth, and you do not have to log in every single time.

For Onepage, that means you can build landing pages and complete funnels directly from Claude, Codex, and similar tools.

You say what you want in chat, and the page appears in your Onepage account.

Long-time readers know this pattern from us: our own servers are already connected to Claude, so the machine can access our websites and make changes directly.

But that setup is usually technical. With Onepage, this works out of the box.

And as always with Onepage: German company, GDPR-compliant, hosted in Germany.

Now comes the part that is the real game changer for us.

With most AI website builders, you get stuck once the machine is done. If you want to change a small thing afterward, you have to prompt again, explain your intent for the tenth time, and hope the AI does not break something else.

Not with Onepage. Everything Claude or the Onepage AI builds can still be edited by non-technical users in the visual editor. Click in, change text, recolor a button, swap an image, done. You are not at the mercy of the machine. You take the wheel whenever you want.

That is why Onepage is currently one of our favorite tools for landing pages and funnels for non-technical people. And MCP lifts it one level higher.

What does that look like in practice?

Here is the exact prompt we sent Claude:

Can you build an alternative newsletter homepage? You know our ainauten.com. Take a look and build a cool alternative in Onepage. Preferably inside the consulting project.

That was it.

A few minutes later, a finished branded landing page was sitting in our Onepage account. Hero, claim, email field, numbers section, everything in the AInauten look. And anything we did not like could be changed with two clicks directly in the section, as shown on the left in the screenshot.

Why it works so quickly for us: Claude knows almost everything about us through our Second Brain. Our projects, our brand, our tone.

We did not have to explain what AInauten looks or sounds like. The machine already knew and poured it straight into Onepage through the MCP.

That is the real leverage in MCPs: your AI knows your context, and through MCP it can immediately turn that context into work inside your tools. From thought to finished page, without detour.

How to set up the Onepage MCP in two minutes:

In Claude, add a new connector and use this URL: https://mcp.onepage.io/

Then confirm once with your Onepage account that Claude is allowed to access it.

That's it. From then on, ask Claude: "What can you do for me in Onepage?" and get started. This does not only work with Claude, but with every AI tool that supports MCPs.

One detail we really liked: MCP usage does not count against the AI credits you otherwise buy from Onepage. That means you can work with AI without burning through your credit budget.

So if you build landing pages or funnels and you are not technical, this is currently the smartest path we know. You can try the whole thing for free here.

🏛️ Harari: AIs Are Born Bureaucrats

Let's close with one short thought experiment.

Yuval Noah Harari gave a talk at Oxford a few weeks ago that is now making the rounds online. The core thesis is not new; he has been telling versions of it for years. But rarely this sharply.

His logic: human power is based on mass cooperation. And mass cooperation runs through bureaucracy: law, finance, administration. All built from words.

And here is the point: AIs are born bureaucrats. No lawyer can memorize every law. An AI can.

That makes them perfect candidates to take over exactly these processes. In Harari's words: AI bankers decide on your loan, AI administrators decide on your university place, AI judges decide your sentence.

You do not have to share Harari's doom tone. We do not fully share it either, although we understand it more and more. The observation that AI first gains real power in invisible administrative processes, not with robots, resonates.

Our pragmatic take is familiar: this is exactly why it is worth understanding these systems now. If you can operate them, you do not just experience them passively.

If you have an hour this weekend, the talk is worth it.

That is it for today. See you in the next issue.

Thanks for being here. Until next week.

Reto & Fabian from AInauten

Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

This issue is for everyone who felt their pulse jump when Fable 5 came back, then almost pasted the same old monster prompt into it. Please do not.

The smartest model does not need a novel. It needs a clear assignment, clear boundaries, and criteria for what "done" means. You will find all of that in our master prompt.

We also show you which tasks truly deserve Fable, where free AI APIs are enough, and why NotebookLM now turns your own sources into Shorts.

Here is what we have for you today:

• 🧠 Our Fable 5 Master Prompt: 5 Tips for Better Output

• 🆓 Free AI APIs: The Easiest Way to Use Them

• 🤳 NotebookLM Now Creates Cool Video Shorts

Let's go.

🧠 Our Fable 5 Master Prompt: 5 Tips for Better Output

Fable 5 is back. If you missed the drama: first we had the hype around Claude Fable as the most powerful AI model, then the reality check when Fable disappeared again and open models suddenly looked like basic equipment.

The lesson back then: do not depend blindly on a single frontier model.

Now Fable is back. And on X, the usual thing is happening: prompt lists, 46-hour demos, cost hacks, agent stacks, and fear of missing out.

You Have Until July 7 to Make the Most of Fable in Your Subscription

For now, Fable is only included in Pro/Max/Team until July 7, and you can use up to 50 percent of your weekly limit on it. After that, it moves to usage credits, meaning pay-as-you-go on top of the subscription.

Anthropic says Fable should return to subscriptions once there is enough capacity. When that happens is still open.

So the practical question is: how do you use Fable sensibly right now without burning your token budget?

Twitter tweet

Tip 1: Prompting: Less Is More

We have learned: the more precise the prompt, the better the result. With Fable, that flips.

The most important basic rule is this: Fable 5 needs less classic prompting.

Old detail prompts can even get in the way, Anthropic says. They look diligent. In reality, they often take away the room where Fable is most useful.

Which makes sense when you look at it: you do not want to spoon-feed the smartest model every sentence. You want to give it a clean playing field.

Goal. Boundaries. Sources. Stop rules. Evidence. Let it figure out the rest along the way.

Or put differently: you do not need to prescribe every intermediate step. You mainly need to say how a good result will be recognized.

Tip 2: First Find Work That Deserves Fable

Fable is expensive, slow, and strong. So its use has to be deliberate, otherwise you are firing a cannon at a tiny problem.

"Make me a marketing strategy" is short and punchy. But short is not automatically good.

A good Fable assignment goes further. It includes context, sources, open decisions, and clear goals with quality criteria. You want to be able to tell whether the work it delivers is actually good.

Every built a useful filter prompt for this: do not let Fable start working immediately. First let it find the right job candidates. Here is our compact everyday version:

Do not start working yet. First check whether this task deserves Fable:
- Does it need multiple sources?
- Are there open decisions or real judgment involved?
- Can we verify at the end whether the result is good?
- Is there a clear finish line?

If too little of that applies, tell me.
Suggest the cheaper path (Sonnet, Opus, human) and stop.
If only parts deserve Fable, mark exactly those parts.

The first cost lever sits before the run. It is called task selection.

The OCR trick, where you push context as an image so Fable can read it more cheaply, is neat as long as it still works. For normal work, the better question is simpler: does Fable need to do this at all?

Tip 3: Let Fable Find Your Gaps

With Fable, the bottleneck is often not the model. It is what you did not say because you were not clear on it yourself yet.

That is why Fable should not start building immediately. It should find the gaps.

Your unknowns come in several kinds: what you said, what you have not decided yet, what you are probably assuming without writing it down, and what you do not even have on your radar.

That is exactly why Fable should first find your blind spots.

Before you work, find my unknowns:
- What did I state clearly?
- What have I not decided yet?
- What am I probably assuming without writing it down?
- Which question would strongly change the result?

Ask me at most five questions.
Start with the question that has the biggest impact on architecture, effort, or result.

If you cannot describe what you want, give references. Old files, screenshots, a good example, a bad example, source code, customer emails.

Anything, as long as it is relevant context.

Fable is good at pulling patterns out of material.

It gets harder when it has to guess your gut feeling.

Tip 4: Use Loops With Evidence and Clear Boundaries

Loop engineering is the next thing everyone is talking about. For everyday work, it does not need to be as big as it sounds.

At its core, a loop is just this: a small next step, a stop rule, evidence, and a decision for the next step. That is the difference between "Fable, do it" and real delegation.

Work in small, evidence-backed steps.
Stop when the next step would only be guessing.

After every loop, deliver:
1. What you did.
2. What evidence supports it.
3. Which unknowns remain open.
4. Whether you would continue, stop, or ask me.

Do not mark anything as done without naming the evidence.
Evidence can be: checked file, test run, screenshot, comparison, source, list of changed places, or clearly named residual risk.
If you could not verify something, write: "Not verified:" and say why.

Fable may run for a long time. But it must not drift endlessly and burn your token budget.

Tip 5: Use Fable Only When It Really Matters

Fable must not become your default model. That is the most important cost tip.

It is the "God mode" for tasks with planning, multiple sources, judgment, long execution, or hard verification.

Tip: inside Claude chats, you can also switch between models.

If Fable itself notices that a part would be cheaper with Sonnet, Opus, or a human, it should say so and not keep calculating out of politeness.

Put this into the assignment:

Use Fable only for the parts that truly need planning, multiple sources, judgment, or long execution.
Delegate simple sub-tasks to other models that can handle them more cheaply.

That creates a practical gain: Fable does not do everything. Fable helps you decide what Fable is worth in the first place.

The new Anthropic Economic Index for June 2026 confirms that 54 percent of Claude Code sessions run on expensive Opus. In chat, it is only 10 percent.

Meaning: autonomy and token use rise together. Better models also like using their peers, even when a simpler, cheaper model would be enough.

The better pattern is: route simple work downward, thinking work upward.

And last but not least: set Fable Effort to high as the default. Use xhigh only for the hardest tasks, and medium/low for routine. Lower effort on Fable often beats xhigh on older models.

Our Take: Fable Needs Less Prompt, but More Frame

Fable 5 is not hungry for your elaborate prompts. It is a very powerful delegation model that needs a clear assignment with boundaries.

If your working mode is sloppy, Fable makes it more expensive. If your working mode is clean, Fable can carry tasks that used to be too large for a chat.

Use it to solve one real knot in your workday: too many tabs, too many notes, a half-described automation, a backlog with no order, a project that has been stuck in your head for weeks.

Write a handoff to Fable that someone on your team would understand.

Here is the Fable 5 Master Prompt that connects all of these tips.

🆓 Free AI APIs: The Easiest Way to Use Them

If you have only used AI in the chat window so far, an API is the next lever. It lets you put models into your own workflows: agents, small scripts, tests, content checks, coding helpers.

Here are the most important highlights:

• OpenRouter is the easiest entry point if you want to test many models with one key. The Free Models Router currently bundles free models. The advantage: OpenAI-compatible API, often just a base URL switch.

• Google AI Studio is the best free path for Gemini models. Strong for long context, multimodality, and quick experiments. Note: free-tier data may be used to improve the product.

• NVIDIA NIM and the API Catalog fit when you want to test models serverlessly. NVIDIA describes free access for development and prototyping.

• OpenAI is not a classic free playground, but there are two paths: Researcher Access with possible API credits and Data-Sharing Tokens for organizations that share traffic with OpenAI.

• Also worth mentioning: Cloudflare Workers AI is good for small automations. The Free Plan includes 10,000 Neurons per day at no extra cost, and text and embedding models offer OpenAI-compatible endpoints. We love Cloudflare and use it for most of our apps.

• You will find more tips in Cheahjs Free LLM API Resources and freellm.net.

Pick one provider and test it with a real task and several models. The clearer the assignment, the better the results.

Our rule of thumb: free APIs are for prep work. Summaries, sorting, first drafts, log explanations. Anything with customer data, code merges, or legal risk does not belong there.

P.S. Here is a starter prompt so you can use this for your own work.

I read this post about free AI APIs:

[paste the post above here]

Based on everything you know about me and my current projects:

1. Where could I use free or cheap AI APIs sensibly?
2. Which 5 tasks are good first tests?
3. Which tasks should I avoid because of privacy, API keys, or legal risk?
4. Which provider from the post fits which test best?
5. What is my next concrete step?

Give me a table with: task, provider, risk, reason, next step.

If you know too little about my projects, ask me a few questions first.

🤳 NotebookLM Now Creates Cool Video Shorts

We have loved Google's NotebookLM from the beginning. After audio, infographics, and video overviews, there are now Short Video Overviews: about 60 seconds, vertical, based on your own sources.

According to Google Help, the format currently runs in English and is first available for paying Google AI Pro/Ultra users, with free availability coming later.

Twitter tweet

Some people call it "TikTok for PDFs". Learning material, just snackable and scrollable.

We already covered NotebookLM as a learning machine in a German Deep Dive, as a graphics hack, and as part of Google's larger source-content studio.

Language note: the older NotebookLM Deep Dive is available only in German, so we did not link it in this English issue.

The new part is this: the output format now fits Shorts, Reels, and quick community explanations directly. For courses, newsletter recaps, or Skool posts, that is pretty obvious.

We wanted to test it right away with our Fable 5 piece above. Not bad. The prompt is in the video description.

Done. See you in the next issue.

Reto & Fabian from AInauten

AIHOY and happy weekend, AInauts!

Maybe you did not catch every useful AI news item, tool, and workflow from the past week. Or maybe you just joined us. No problem.

We pulled together our highlights from the week for you:

• 🐳 7 Token-Maxxing rules: how to get more from your AI subscription

• 🤖 Claude @: why the next agent sits in your team chat

• 🥲 GPT-5.6 is here. The real question is when you get to use it

• 🎨 Great AI videos, avatars, and more with this tool

• 🔮 A cool way to make better decisions with AI

• 😿 Is the golden AI window closing for consultants and agencies?

At the end, you will also find the most important Quick News, so you can scan the relevant developments in one go. Ready? Let's go.

🐳 7 Token-Maxxing rules: how to get more from your AI subscription

AI feels limitless until your subscription runs out right when the agent finally understands the job. The problem is rarely one huge prompt. More often, your setup slowly burns through context: old chats, uploads, enabled tools, correction loops, and premium models doing tiny tasks.

These 7 rules are not a penny-pinching list. They are better leadership. Separate planning from execution. Give a goal instead of a vague request. Keep context small. Turn files into the working surface. Start fresh earlier. Load tools only when they have a real job.

Token-Maxxing is the new work discipline for power users. If you dump everything into one chat, you pay twice: first with limits, then with worse answers.

👉 Read the 7 token rules »

How Much Is Your Billing Lag Actually Costing You?

Most SaaS finance teams know their billing process isn't perfect. Few know what it's actually costing them.

Answer 5 quick questions — contracts signed per month, ACV, days to first invoice, error rate, DSO — and the Tabs Billing Lag Calculator gives you a dollar figure benchmarked against top SaaS companies.

It takes two minutes. The number might surprise you.

Calculate your billing lag and see where you stand.

Calculate your billing lag

🤖 Claude @: why the next agent sits in your team chat

Claude is moving from the solo chat into the team room. With Claude @, you mention the agent in a channel, give it context from the discussion, and let it keep working asynchronously. This is more than Slack convenience: the agent does not just get a task, it gets the social work context around that task.

That is exactly where things become useful and risky. Team chats contain decisions, half-knowledge, politics, customer context, and old promises. If Claude can read along, it can help better. But it can also misunderstand far more when permissions and boundaries are fuzzy.

Team agents will come because they are useful. But the first good use case is not "read everything." It is one clearly bounded channel with one recurring job. Otherwise your team does not get an assistant. It gets a very polite risk.

👉 Read the Claude team-chat take »

🥲 GPT-5.6 is here. The real question is when you get to use it

OpenAI has previewed GPT-5.6 Sol, and as usual, the model is only half the story. Stronger, safer, more cyber protection, more red-teaming, the next frontier tier. But as users, we mainly want to know: when does it land in our subscription?

After the Claude Fable 5 mess, we know that access is political, economic, and technical. The best model does not help if it is blocked, throttled, or hidden behind expensive credits at the exact moment you need it.

Our recommendation: do not build your work around one model. Build it around clean tasks, local files, swappable model tiers, and clear handoffs. Then GPT-5.6 becomes an upgrade instead of a nervous breakdown.

👉 Read the GPT-5.6 take »

🎨 Great AI videos, avatars, and more with this tool

We are having a lot of fun with AI video right now. Not just "prompt in, clip out," but campaign, reference images, character, product shots, script, voiceover, subtitles, and eventually a usable ad spot.

That is where Higgsfield is currently strong. The "Supercomputer" tool is basically a creative agent: you give it a brief, it plans subtasks, chooses models, and builds assets. For creators, shops, and agencies, this suddenly feels much closer to real production than to a toy. It is not perfect, it is not cheap, and German is still shaky.

So Higgsfield is not a magic advertising-film machine yet. But it clearly shows where creative work is going: less single tool, more production system. If you test products, avatars, or UGC, this platform is worth a serious try.

👉 Read the Higgsfield workflow »

🔮 A cool way to make better decisions with AI

If you ask your AI whether your plan is good, you often get ten reasons why it will be not just good, but brilliant. Nice for the ego, bad for business. So we flip the question: premortem instead of applause.

The prompt is simple: It is one year later, the plan has failed, tell me how it died. Suddenly the AI is not looking for confirmation. It looks for breakpoints, blind spots, early warning signs, and kill criteria. It is an old and effective decision technique, neatly wrapped into agent form.

👉 Read the premortem skill »

😿 Is the golden AI window closing for consultants and agencies?

The last few months were comfortable for many service providers: AI makes the work faster, while clients still pay for old hour counts. That deal is starting to crack. Once clients realize that research, analysis, concepts, and slides can move faster, the billable hour suddenly becomes the wrong metric.

Large consultancies already feel this and are testing more fixed prices, outcome logic, and performance-based models. Smaller agencies will face the same shift later and harder. If you only sell "we use AI," you will soon have no argument left. The client can do that too. Sell outcome, judgment, execution, and risk ownership, not the time your agent just saved.

👉 Read the agency and consulting take »

AI News Quickie: the HAI-lights from the industry

AI never sleeps. This week was all about access and control: OpenAI is flirting with sovereign involvement, Fable 5 is back, Microsoft is training agent skills, AWS is putting models into regulated clouds, and even hotel reviews show why AI summaries are not automatically truth.

🎯 OpenAI is pushing on several fronts

• Sam Altman reportedly proposed giving 5 percent of OpenAI to a US sovereign wealth fund.

• OpenAI uses ChatGPT adoption data to show how usage is expanding across regions and user groups.

• With GeneBench-Pro, OpenAI is testing models on more realistic computational biology tasks.

• GPT-5.6 Sol is being positioned as the next model generation, with stronger cyber focus and tougher safeguards.

• OpenAI has announced DevDay 2026 for September 29. For builders, that means: do not reinvent your API roadmap in September; keep workflows modular now.

• The Economic Research Exchange is accepting applications until July 5. OpenAI wants evidence on economic impact because gut feeling about AI jobs is becoming too expensive.

• AWS is bringing OpenAI gpt-oss and NVIDIA Nemotron to GovCloud. Regulated organizations get open models hosted in a compliant environment.

• OpenAI is teasing Codex hardware with Work Louder. Maybe it is a niche gadget, maybe the start of agent buttons instead of app icons.

🏛️ Claude Fable is back

• Anthropic is turning Fable 5 back on after export controls were lifted. The whole incident shows that model access can always be temporary.

• The Fable 5 dispute is not gone, according to this Guardian readback; it has just been repackaged. Security review sounds better than a blanket stop, but it is still political access control.

• Fable 5 was reportedly significantly faster than Opus 4.8 in a coding test.

• Claude Sonnet 5 is available on AWS and aims for Opus-like capability at Sonnet pricing. For many teams, this may become the new cheaper default.

• Anthropic is launching Claude Science as a workbench for scientists.

• Claude Remote Routines are being described as a work ecosystem: Chat, Code, Cowork, Design, and Routines. The model is becoming a workspace.

• A Mythos 6 leak allegedly promises autonomous vulnerability discovery.

🖥️ Google, Microsoft, and AWS are building the agent workbench

• Google summarizes its AI updates from June, from Gemini to Android. The common thread: AI is becoming more of an operating layer.

• NotebookLM is getting vertical 60-second clips, useful for snackable social media content.

• Google's new Home Speaker shows in an early review: the hardware is good, but Gemini as a household assistant still has room to grow.

• Microsoft Research introduces SkillOpt, a way to optimize agent skills like trainable parameters, and shows Memora, a memory representation that can sharply reduce context tokens.

• AWS describes resilience patterns for Bedrock and pairs Nova 2 Lite with Claude for cheaper document processing. Token-Maxxing for enterprise.

🏗️ Infrastructure, money, and power are getting tougher

• NVIDIA is expanding the AI factory model with Capital Partners. Jensen is no longer just selling chips, but financing, cloud logic, and revenue participation.

• Meta apparently wants to turn excess compute into a cloud product and rent out GPU overcapacity.

• The UN warns that AI can worsen global inequality.

• Local AI is becoming more attractive as protection against expensive cloud dependency.

• Zurich is emerging as a quiet R&D hotspot for the global AI elite. Research follows talent, tax logic, calm, and proximity to major companies.

• San Francisco's AI boom makes even $180,000 salaries look small.

• Bhavin Turakhia is putting $30 million into Neo, an AI-native Office alternative.

• AI stocks are losing ground, according to the Guardian, but the market is still far from a major crash.

🛡️ Society, tools, and the real workplace

• Erin Brockovich is taking aim at AI data centers. AI infrastructure is local: water, electricity, land, resistance.

• Tripadvisor's AI reportedly softened serious hotel complaints. That is a useful reminder: AI summaries need scrutiny.

• A CEO reportedly threatened termination over unreviewed ChatGPT emails. Translation: AI slop is not productivity.

• Ford is bringing experienced engineers back because AI cameras missed quality risks.

• The Commonwealth Short Story Prize honored a controversial story despite AI-use claims, while e-book lending app Libby is getting an AI filter.

• Tidal is cutting royalties for fully AI-generated music.

• People who frequently use AI for health topics are more likely to believe vaccine myths, according to a KFF survey.

• A US bill aims to ban the sale of health and location data from AI chats.

• Rogbid is selling the VisionPro as a $120 AI glasses product with camera and live translation.

• Formula 1 teams are wrestling with how to integrate AI sensibly without hollowing out the sport. Better simulation, yes. Invisible autopilot, no.

• SpaceX reportedly showed an AI handset prototype; Musk denies it. Everyone is searching for the next interface after the smartphone.

Our thread of the week: AI is growing up, but it is not automatically trustworthy. Whoever manages access, context, cost, and quality well wins. Whoever only chases the strongest model gets stuck again at the next limit.

Done. But no reason to be sad. The AInauts will be back soon with fresh material for you.

See you Monday with a new round of news, workflows, and insights.

Your AInauts
Fabian & Reto

P.S.: Follow us on the socials! X, LinkedIn, Facebook, Insta, YouTube, TikTok

AIHOY AInauten,

Welcome to the latest edition of your favorite newsletter.

Quickly, two Anthropic notes before we leave you alone with it: by the time you read this, Fable 5 should be available again. And Claude Sonnet 5 is here, landing almost at Opus level for knowledge work.

What that means for you: for most everyday work with Claude Cowork and similar setups, Sonnet is now enough. We tested it, it works really well, and if you are paying only around 20 euros a month, this should let you do noticeably more with Claude.

Now let’s get into today’s topics. Lots of practical things to try, a free skill on top, and one bigger industry question at the end.

Here is what we have for you today:

• 🎨 Great AI Videos, Avatars, and More With This Tool

• 🔮 A Cool Way to Make Better Decisions With AI (+ Free Skill)

• 😿 Is the Golden AI Window Closing for Consultants and Agencies?

Let’s go.

Not just another AI newsletter.

Not just another AI newsletter. MavSource aggregates updates from all major AI newsletters, podcasts, company news, AI labs, and hundreds of other sources — then summarizes what matters, analyzes emerging trends, and adds founder commentary. One 5-minute daily email. Free.

Subscribe For Free Now

🎨 Great AI Videos, Avatars, and More With This Tool

This week, we are deep in AI video again.

So much is happening here right now that it is worth taking another look and testing a few things yourself.

Last week we went deep into Higgsfield. It is one of the most powerful video tools on the market because it is not just one model. It is a whole platform that gives you access to almost all relevant video models.

The team behind it is fast, agile, and creative. New features keep landing, and the point here is simple: this is worth testing if you do anything with creative production.

The Supercomputer: a Whole Creative Team in Chat

The fastest way to understand how powerful Higgsfield can be is one feature: the Higgsfield Supercomputer.

It is basically an ultimate creative agent. You give it a simple prompt, and it can develop the whole thing from A to Z.

You type a normal-language briefing, for example: “Create a campaign for product X.”

The Supercomputer plans the steps, chooses the right models, runs several subtasks in parallel, and delivers finished assets.

In our test, we asked it to create a 30-second UGC spot for a vitamin C serum, first generate three product reference images, create a UGC character named Sophie, and turn everything into a TikTok-style video with voiceover and subtitles.

The Supercomputer then created the assets: character, product, product shots, UGC script, and the video itself.

First attempt, impressive quality. Almost too easy.

That shows what Higgsfield can do. Product shoots, online-shop visuals, consistent AI avatars, UGC videos with the same character, all of that is possible there.

One last strong point: Higgsfield has an MCP and CLI, so you can connect it to Claude, Codex, and similar tools and let your agents create assets for you.

How to Start and Our Take

The best place to start is the Higgsfield Supercomputer. Create an account, choose a plan, and play with it in chat. Thanks to the new Sonnet 5 model, it is faster and cheaper too.

Important downsides:

• Everything video-related is still expensive. Fifty euros a month can disappear quickly.

• German speech in videos is still hit or miss.

• English is excellent. Hopefully German catches up soon.

Still, it is a very strong tool with a huge number of creative templates and powerful features.

Next week we will continue with more ways to use AI video. We are learning a lot ourselves right now and will share more soon.

Is Your Finance Stack Ready for Usage-Based Revenue?

Usage-based and hybrid pricing models are reshaping B2B revenue—but for finance teams, they're also introducing serious complexity around revenue recognition, forecasting, and operations.

Join Rebecca Schwartz, Co-founder of Tabs, and Amit Dhir, Partner at PwC, for a live session on June 10th from 1–2PM EDT. They'll break down how modern pricing decisions ripple through your financial workflows—and how to scale without adding manual overhead.

You'll leave with practical frameworks, real-world examples, and a clear path forward.

Register now to secure your spot—and if you can't make it live, a recording will be available.

Can't wait? Explore more Tabs webinars to catch up on past sessions.

Save My Spot

🔮 A Cool Way to Make Better Decisions With AI (+ Free Skill)

If you have been reading us for a while, you know we love using AI for strategy work and as a sparring partner.

We like taking concepts that used to require expensive coaches, trainers, or a full-day workshop and turning them into practical AI workflows.

Today’s setup is this: you have a plan, you ask your AI, “Is this good?”, and it finds ten reasons why it will be great.

You feel validated, start executing, and a few weeks later exactly the obvious thing blows up in your face.

That happens because language models tend to agree by default. They want to make you happy.

Today’s practical hack gets around that. We use an old proven technique: the premortem.

Instead of asking afterward, “What went wrong?”, you fast-forward: It is one year later. The plan failed. Tell me how it died.

That one sentence changes the job. AI stops polishing your plan and starts describing every path where it can fall apart: failure story, hidden assumption, early warning signs, and kill criteria.

This is not a toy method. It comes from Gary Klein and was published in Harvard Business Review. Daniel Kahneman called it one of his most valuable decision techniques.

We have used it for two strategic AInauten decisions, and it showed us blind spots we would otherwise have missed.

How It Works

We prepared a full setup so you can use it immediately: our version is a bit sharper than the ones currently floating around, with a “what did we miss?” countercheck and clear kill criteria.

You can download it here:

→ Get the premortem for free

The download includes two versions: one as a skill for Claude Code and similar agents that builds a report file, and one you can paste into chatbots. Pick the one that fits your workflow.

These skills are especially useful if you already have a second brain, because the machine can include your real context.

Have fun with it, and send us your experiences.

😿 Is the Golden AI Window Closing for Consultants and Agencies?

Let’s end with a topic that many consultants, service providers, and agencies are wrestling with right now.

The old service-business deal was simple: smart people spend hours on hard problems, and the client pays for those hours.

For many, the last few months were golden: AI tools around Claude Cowork and Codex became useful, work got done much faster, but clients were still willing to pay old hour volumes.

That is changing.

The Wall Street Journal reports that large consultancies want to move away from hourly billing because AI makes a lot of work faster and cheaper.

No, they do not really want to. They have to. Clients know that work got easier, so they want it cheaper.

The billable hour punishes the fast. If you become more efficient, you can bill less value.

Large consultancies are already rebuilding: Deloitte has shown its consultants a curve where classic hourly consulting shrinks sharply by 2035. At McKinsey, more than a quarter of global fees are already tied to outcomes instead of hours.

So What Now?

The short-term path for service providers is probably away from hourly billing and toward fixed-price packages with clear goals and outcomes.

In some cases, 100% performance-based deals can also make sense. In our performance marketing campaigns, for example, we have long worked on CPL and CPA models: clients pay per lead or sale.

That puts us at risk. It works in online marketing, but of course not everywhere.

Our Take

As always, it comes down to incentives.

• The billable hour punishes the fast. If you become more efficient, you can bill less value.

• “We pay per closed ticket” rewards speed and invites quality cuts.

The fact is: the window of great AI arbitrage is closing. Consultants, service providers, and agencies have to adapt.

We think that is positive. If everyone is incentivized to make projects successful, that is healthier. It also gets more complex.

If service providers take on more risk, margins can rise significantly too.

That is it for today. As always, we hope something in here helps you move forward.

See you next week.

Reto & Fabian from AInauten

AIHOY AInauten,

Welcome to the latest edition of your favorite newsletter.

AI feels almost limitless right now. More models, more agents, more tools, more automations. At the same time, new limits are appearing: your context window is full, your subscription is throttled, your favorite model is unavailable, or access is suddenly being negotiated politically.

This issue is about that new scarcity. We show you how to use your token budget better, why Claude in the team chat changes more than Slack convenience, and why GPT-5.6 Sol could be a preview of how frontier models may be distributed from now on. Sounds good?

Here is what we have for you today:

• 🐳 7 Token-Maxxing Rules: How to Get More Out of Your Subscription

• 🤖 Claude @: Why the Next Agent Sits in the Team Chat

• 🥲 GPT-5.6 Is Here. The Question Is When You Get to Use It.

Let’s go.

Keep up with AI in 5 minutes a day

Keep up with AI in 5 minutes a day. MavSource aggregates updates from all major AI newsletters, podcasts, companies, AI labs, and hundreds of other sources — then summarizes the key updates and analyzes the trends shaping AI. Free. Daily.

Get the free brief →

🐳 7 Token-Maxxing Rules: How to Get More Out of Your Subscription

If your AI subscription hits the limit too early, it is rarely because of one big prompt. Most of the time, your setup eats the context window one small bite at a time.

In practice, we see two groups fighting the same problem:

• Power users: if you pay for a large subscription, you want to actually use the token budget you paid for.

• Occasional users: if you are on a free plan or a smaller subscription, you do not want to hit the limit exactly when the AI finally understands what needs to happen.

Your AI limit is rarely consumed by one single chat. It is slowly drained by old threads, huge uploads, unclear tasks, active tools, correction loops, and premium models doing tiny chores.

We pulled together the 7 most important rules so your subscription lasts longer and you get more value from it.

Note: Desktop apps such as Claude Desktop and Codex benefit most from these tips because they can work with local files, folders, and tools. That is the Files-over-Tools principle.

1. Separate Planning and Execution

The mistake is not always the wrong model. Often it is using one model for everything.

If you have a lot of material and first need orientation, a cheaper model is often enough for scouting. It finds levers, sorts options, and evaluates effort.

If the decision itself is hard, flip the setup: let a strong model plan. A cheaper model can then execute the clearly described individual steps.

Short planning prompt:

Separate this task into planning and execution.
Tell me which layer needs a strong model.
Suggest one cheaper and one stronger model tier.
Do not execute until I approve the plan.

Model names change depending on the interface. As a rough guide:

• Instant, Medium, or Sonnet-level models are often enough for scouting, structure, drafts, and routine work.

• High, Extra-High, Opus, or Ultracode-level models belong at points where a wrong decision has consequences.

• Pro Extended or Max modes are not the default gear. Use them when the input is difficult, the error would be expensive, or the result directly creates value.

Model selection in ChatGPT

Model selection in Codex

Model selection in Claude

Rule of thumb: think with the expensive model, execute with the cheaper one. Or scout cheaply, solve expensively.

The main thing: do not blast everything through the same model tier.

2. Give a Clear Goal, Not “Just Make It Better”

Weak assignments cost tokens, time, and nerves because the model first has to guess what matters. We admit it: “make this better” was one of our favorite prompts last year after reviewing our ChatGPT usage. But we learned from it.

So include these four lines and briefly force your own brain to formulate them properly. If you need help, ask the agent to give you suitable answer options for each field.

Goal: [What should get better?]
Material: [Which files or links count?]
Output: [What should exist at the end?]
Boundary: [What should not be touched?]

3. Smaller Context Beats Larger Context

Tokens are the model’s currency. Your question costs tokens. The answer costs tokens. Older messages, files, tools, and outputs cost tokens too.

It can even go so far that the entire chat history is sent along with every interaction and keeps eating tokens.

So ask yourself: what does this chat need before it can work? If the answer is “everything,” it will become slow, expensive, and less precise.

Better: provide only the relevant context, not your entire archive.

4. Turn Files Into the Workspace

Storage is cheap. Context is expensive.

A file on your drive costs no tokens as long as the model does not read it. It becomes expensive only when you paste it into the chat, upload it, or actively ask the model to process it. If you upload it into the chat, it fills the context window.

So: Files over Tools.

Pull your knowledge out of chats and save it locally in Markdown files. You can use the prompt from tip 5 for that.

Desktop and CLI agents such as OpenAI Codex or the Claude Desktop app can work very well with local Markdown files.

5. Start Fresh Earlier or Summarize Your Chat Cleanly

Long chats feel convenient, but at some point they get sluggish.

When one chat contains several topics, old decisions, and correction loops, context usage goes up. At the same time, the model often answers more vaguely.

So make a clean handoff and start a new chat.

We put together this prompt to summarize any chat so you have the relevant context for a fresh start.

That saves tokens and prevents outdated knowledge from slipping in.

6. Load Tools Only When You Need Them

Connectors, MCPs, plugins, and browser tools are powerful. But they cost tokens in the context window.

Put differently: if you only want to write a text, you do not need a whole toolbox with 27 other tools active.

Check the base usage. In Claude Code, /usage shows what an empty chat already consumes. In Codex, /status shows how much context is already occupied.

In ChatGPT on the web and Claude Cowork, you can manually check which tools, plugins, and connectors are active.

The rule stays the same: anything active should have a job. Disable what you do not need.

Advanced users can turn some MCPs into CLIs. Then the tool stays outside the chat and is called only when needed.

Review my active tools, MCPs, plugins, and connectors.
Sort them into: constantly needed, rarely needed, currently unnecessary.
Recommend what I should disable or move into a CLI
so more context remains for the actual task.

For that, look at MCPorter and mcp2cli. We explained the background in our AInauten Knowledge Base: CLI or MCP: when do you use what?

7. AI Is a Slot Machine. Use That.

AI is sometimes a slot machine: one run lands, the next one goes completely sideways.

Use that deliberately. If the model is clearly moving in the wrong direction, do not feed the chat with five repair rounds. That is more context you drag along, and worse, it is flawed context.

Edit a prompt in ChatGPT

Edit a prompt in Codex

Stop early, edit the original prompt, or start a new chat with a clean briefing. That helps you avoid context contamination: wrong assumptions, half-corrections, and failed attempts otherwise remain in the context and pull the next run off track.

You can also try different paths in parallel by using branch or fork features.

Create a branch in ChatGPT

Fork a chat in Claude Desktop

The Best Subscription Hack Is Better Direction

If you constantly hit your limit, memorize these tips. They are your checklist.

If you never hit your limit, they are your quality lever. Give it a try!

Six people doing the work. Your headcount is one.

Your finance close runs in #finance. Stripe and QuickBooks reconciled, runway updated, posted Sunday night without you asking.

Engineering review lands in #eng. Viktor pulled the open PRs, left comments on auth-refactor, flagged a dependency blocking api-pagination.

Campaign brief lands in #growth: Meta CPA up 18%, recommendation to pause broad match, a draft landing page already deployed for the variant test.

You hired him on day zero. He lives in Slack and Microsoft Teams alongside your contractors and investors, connects to 3,000+ tools, pushes back when you ship something dumb.

"Viktor is now an integral team member, and after weeks of use we still feel we haven't uncovered the full potential." Patrick, Director, Yarra Web.

Get Started for Free

🤖 Claude @: Why the Next Agent Sits in the Team Chat

The new Claude Tag feature looks at first like another Slack integration. It is in beta for Claude Team and Enterprise. A team owner or admin sets up Claude for Slack, and after that the agent can be mentioned in approved channels.

But it is more than a simple bot in chat, because something changes with that little @: the private prompt gets an audience.

The Chat Gets Witnesses

Until now, AI work was often invisible. Someone opens a chatbot, writes a prompt with context, has something built, and brings the result back to the team. The output is there. The path remains hidden.

With @Claude, that process suddenly becomes transparent. Prompting gets witnesses. Delegation gets witnesses. Errors get witnesses. The private super-assistant becomes a team member with chat history.

Andrej Karpathy, the former OpenAI co-founder and Tesla AI lead, recently joined Anthropic for pre-training and is already excited about the new teammate.

Twitter tweet

You Are Sharing Your Workspaces With Permanent Agents

This fits a bigger trend: AI is moving out of the chat window and into the surfaces where we already work.

• Gemini in Gmail summarizes threads, suggests replies, and pulls from Drive files.

• Microsoft 365 Copilot sits in Word, Excel, PowerPoint, Outlook, and Teams.

• Meta is rolling out Meta AI in WhatsApp, Instagram, Facebook, and Messenger.

Claude Tag is the team-chat version of that. The agent sits where status, shortcuts, misunderstandings, and office politics already live.

Slack is not a neutral tool. Slack is the office hallway with search. When an agent shows up there, output, visibility, trust, and responsibility all land in the same thread.

In a private chat, Claude belongs to you. In a team channel, Claude belongs to the room.

Three people write along, two have data access, one person can see customer data, and someone else is only a guest. Whose context may Claude use? Whose boundaries apply? Who cleans up when the agent puts something wrong into the thread?

Anthropic addresses this with an agent identity model. Claude gets its own access, memories, and rules per workspace or channel. According to the docs, access follows the channel, not the person asking at that moment.

The new question is: who is allowed to bring AI into which workspace, and who is responsible when that room starts working differently?

Our Take: AI Disappears Into the Interface and Becomes Everywhere

Claude Tag is a strong signal for where this is going: the most useful agent sits where the next action happens.

Chatbots learn you and your company message by message. That is the value. It is convenient. And it is a shift in power.

That is also why some teams will feel weird about it.

🥲 GPT-5.6 Is Here. The Question Is When You Get to Use It.

Last week, it sounded like: GPT-5.6 is coming. Buckle up.

Now GPT-5.6 Sol is here. Just not for us.

Access starts in a limited way after consultation with the U.S. government. No public waitlist. No regular users. Only selected partners get to try it first.

That is more than bureaucracy. Until now, “partners first, everyone else later” was normal release logic. Now the U.S. government is visibly sitting near the lever. OpenAI itself says this process should not become the norm because it keeps the best tools away from users, developers, companies, and cyber defenders.

And OpenAI is not alone. Claude Mythos 5 is again allowed to go to certain U.S. organizations in critical infrastructure and cyber defense after negotiations. Fable 5 remains stuck. This is where the uncomfortable question starts: will regular users even see the strongest models in the future, or only the public, safely limited version?

The meme version: Anthropic CEO Dario Amodei ruined it for everyone. If labs themselves say their models could become dangerous, nobody should be surprised when officials eventually stand near the release button and the newest supermodels stay locked away for state-approved purposes.

But fine. If we do not get the new model, at least we get good AI drama and the hope that Fable 5 comes back before July 1, if you believe the gamblers on Polymarket.

Our Take: Your AI Stack Needs a Plan B

The most important question is not: when do I get GPT-5.6?

We are curious about GPT-5.6 and will test it as soon as we can. Of course. According to benchmarks, it is supposed to be ahead of Fable in some areas.

The more important question is: can my workflow switch if my favorite model is unavailable tomorrow?

So: store context in files and get it out of chatbots. Keep setups model-neutral. Build offline pipelines. See our post on the offline AI setup.

If we do not get a new model, at least we get more clarity about why Sam Altman was fired from OpenAI and how he ended up back in the CEO chair. “You should, unfortunately, be worried about Sam Altman” is a very worthwhile film with a lot of background.

Watch the video here

Done. See you in the next issue.

Reto & Fabian from AInauten

AIHOY and happy weekend, AInauts!

Maybe you missed a few of last week's most useful AI news, tools, and workflow ideas. No stress.

Here are our highlights from the past week:

• 😵 SaaS Killer: Your Own App Instead of Another Subscription

• 👨‍💻 ChatGPT Is Moving From Answer Box to Workspace

• 🔎 Germany's AI Location Index 2026

• 🛒 4 E-Commerce AI Business Ideas Without Your Own Product

• 🪄 Onepage AI 2.0: Vibe Coding Meets No-Code

• 🐡 Fugu: Is This the New Model Approach?

And at the end, you get the most important Quick News, so you can catch up on the relevant shifts in one read. Ready? Let's go!

Investors see ANOTHER return from Masterworks (!!!!)

That’s 6 sales in 7 months. 29 all time. And the performance?

16.5%, 17.6%, and 17.8%, net annualized returns on sold works held longer than one year (See all 29 at Masterworks.com)

It’s not from stocks, private equity, or real estate… it’s from contemporary and post war art. Crazy, right?

With Masterworks, you don’t need to be a BILLIONAIRE to invest in multi-million dollar art anymore.

Historically, the segment overall has had attractive appreciation and low correlation to stocks.*

Masterworks targets works featuring legends like Banksy, Basquiat, and Picasso, identifying what they believe to have significant long-term appreciation potential, not just at the artist level but at the level of individual artworks.

As one of the largest players in the art market, with $1.3 billion invested over 500 artworks, they pass critical advantages through to their 70,000+ members to add art to their portfolios strategically.

Looking to diversify your investments in 2026?

Click here to skip the waitlist

^{*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.}

😵 SaaS Killer: Your Own App Instead of Another Subscription

The best AI moment of the week did not come from a lab. It came from the community. Ralf did not subscribe to a meeting-notes tool. He built a small desktop app that transcribes audio, separates speakers, and gets the job done for cents.

This is not a "everyone builds a SaaS now" story. The more useful point is smaller: when a subscription annoys you and the workflow is clear, an agent can often build a good shortcut. Maybe not perfect. But often good enough to save money, friction, and tool fatigue.

The new AI power user does not buy software on reflex. They describe the job, let an agent plan and build, then test with real files. After that, you know whether you can cancel the subscription or finally pay for it with a good reason.

👉 Read the SaaS killer prompt »

👨‍💻 ChatGPT Is Moving From Answer Box to Workspace

OpenAI is moving ChatGPT and Codex closer to your actual work. Scheduled Tasks, Computer Use, Chrome context, a dedicated browser, Memories, and Chronicle all point in the same direction: less "ask me again later," more "I'll take care of it."

That is useful, and sensitive. If an agent can work in your browser, it can see logged-in systems. If it gets Computer Use, it can click and type. That is a productive teammate with access. Good delegation starts with boundaries, not blind trust.

👉 Read about the new workspace »

🔎 Germany's AI Location Index 2026

Berlin leads the new German AI location index ahead of Munich and Stuttgart. The ranking looks at jobs, research, education, startup ecosystem, and salary after rent. Sounds dry, but it is a useful reality check.

An AI location is not just a city with a few loud events. A place becomes relevant when education, capital, demand, research, and companies become visible at the same time.

That is why the gap behind the top three is so interesting. Germany often debates big AI strategies. Daily progress happens in cities, teams, and job postings.

👉 Read the location reality check »

🛒 4 E-Commerce AI Business Ideas Without Your Own Product

Shopify has opened a door with the Universal Commerce Protocol that is especially interesting for small builders. AI agents can find products in Shopify catalogs, build carts, and prepare purchases. You do not necessarily need your own shop.

The useful part is not only the API. The useful part is the arbitrage: if you understand a niche, you can build shopping advice, gift concierge workflows, best-value finders, or shopping inside your own content.

Not everyone will turn that into a lasting business. But if you have an audience, niche, or community, you can test much faster than before.

👉 Read the 4 e-commerce ideas »

🪄 Onepage AI 2.0: Vibe Coding Meets No-Code

Building websites with AI has become easy. Changing them later without drama is still the problem. That is where Onepage AI 2.0 gets interesting.

The smart part is Promptable Controls. You can tell the AI which controls you need, such as accent color, headline, or button text, and get an editing panel directly on the section.

That is much closer to real work than prompting "change the third paragraph" five times. We find this combination useful for non-techies because it does not stop at the first pretty draft.

👉 Read about Onepage AI 2.0 »

🐡 Fugu: Is This the New Model Approach?

After OpenRouter Fusion, Sakana's Fugu is the next attempt to get top performance not from one single model, but from a team of models. An orchestrator plans, distributes, checks, and combines the answer.

The principle is strong. The current day-to-day experience is less glamorous. Our first tests: still buggy, not clearly better than Opus and friends, and again with availability catches for DACH users. The direction still matters. AI models will work less like individual geniuses and more like small teams.

👉 Read the Fugu take »

📰 AI News Quickie: The Highlights From the Industry

AI never sleeps, even if this week was a little quieter. Here are the highlights.

🎯 OpenAI Builds Out the Enterprise Stack

• OpenAI started a limited preview of the new model line with GPT-5.6 Sol.

• OpenAI showed in a new study how agents like Codex take on longer and more complex tasks. That fits the week pretty well: AI is sliding from chat into real work.

• Scheduled Tasks in ChatGPT now get their own task interface. Small routines move out of your head and into the chat.

• OpenAI launched Patch the Planet, a program for open-source maintainers who want to fix security vulnerabilities with AI support.

• The Daybreak initiative is getting tools to find and patch vulnerabilities at large scale. Security is becoming more of an agent job.

• ChatGPT's market share has fallen below 50 percent. ChatGPT is still huge, but the assistant market is spreading out.

• OpenAI is supporting new standards for advanced AI.

• OpenAI and Broadcom introduced Jalapeño, their own inference chip. Whoever controls the models also wants to control the machine underneath.

• Samsung is rolling out ChatGPT Enterprise and Codex broadly across its Device eXperience division. AI becomes operating infrastructure inside the company.

• OpenAI says GPT-5 helped with an immunology puzzle that had been open for years.

• OpenAI is making frontier models and Codex available on AWS, going where companies already run infrastructure.

• Getty Images is allowing OpenAI to show licensed content in ChatGPT Search and discovery.

🏛️ Claude, Anthropic, and Agents in Teams

• Anthropic launched Claude Tag for Slack. You mention Claude in a channel, the agent picks up context from the team room, and continues asynchronously.

• The full Claude Desktop experience is coming to enterprise clouds from AWS, Google Cloud, and Microsoft Foundry. For IT teams, that means more agent work inside managed infrastructure.

• Claude Code Artifacts are becoming shareable live dashboards. Prompt outputs turn into small team workspaces.

• Microsoft's Copilot Cowork is generally available. Office work gets agents with cost centers, admin boundaries, and company context.

• The dispute around Anthropic export controls shows why cyber rules fit AI models badly.

🖥️ Google, Design, and New Interfaces

• Google is building Computer Use into Gemini 3.5 Flash. The browser is no longer just read; it can finally be operated.

• Gemini in Chrome will be able to use Select from screen to understand selected content directly. Small interface change, big everyday effect.

• At Config, Figma introduced new motion, shader, and code tools. Design teams get less blank canvas and more prompt plus control.

• Meta and EssilorLuxottica launched Meta Glasses starting at $299, this time with everyday usefulness as the promise.

• ByteDance introduced Seedance 2.5, a video model for 30-second clips from prompts.

• Midjourney announced Midjourney Medical, a wild step toward ultrasound scanners.

• xAI pushed Grok Imagine Video 1.5 to the top of video leaderboards.

• Google Finance is getting AI briefings, portfolio uploads, and an Android app. You describe your market briefing once, and Google delivers it regularly.

🤖 Models, Chips, and Infrastructure

• Sakana AI launched Fugu and Fugu Ultra, a multi-agent system behind an OpenAI-compatible API.

• GLM-5.2 is one of the strongest open-weights text models. Open models are moving closer to frontier workflows.

• Groq confirmed a $650 million round and is rebuilding the team after NVIDIA's talent move.

• NVIDIA launched the BioNeMo Agent Toolkit. Life sciences teams get agent tools, not just model access.

• Mistral released OCR 4 for enterprise documents in 170 languages. Not flashy, but important: many company problems still live in PDFs and scans.

• China is planning a national AI data center network with domestic chips. Sovereignty is being built here, not just debated.

• Google is investing in Alabama infrastructure. AI compute is becoming a regional economic factor and an energy question at the same time.

🛡️ Security, Law, and Market

• RAISE US launched with more than $500 million for AI reskilling.

• AI companies are buying up old books and destroying them. Data hunger can look very analog.

• A Dialog data leak made a secret tech and power elite visible. Privacy failures do not only hit normal users.

• A lawsuit against Kalibrate accuses AI-powered pricing software of coordinating gas station prices. Algorithmic prices are not automatically neutral.

• A backdoor into the living room? Smart devices remain a security problem, and AI makes the analysis easier, not the devices safer.

Our thread of the week: AI is moving out of the chat window and into real workspaces. That is useful when you delegate small jobs cleanly. It is dangerous when you hand out access before the workflow has boundaries.

That's it for this week.

The AInauten will be back soon with fresh AI food for you. See you Monday with a new round of news, hacks, and insights.

Your AInauten
Fabian & Reto

P.S.: Follow us on social: X, LinkedIn, Facebook, Instagram, YouTube, TikTok

Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

Most likely, right after we send today's newsletter, the new GPT-5.6 model will drop. Probably with a two-million-token context window and brand-new business consulting skills.

And maybe a drone mode. We are exaggerating. A little. But after the past few model weeks, nothing would shock us anymore.

We will touch on new model approaches towards the end. Before that, today's issue is practical: a few business ideas for the AI age and simple ways to build websites.

Here is what we have in store for you today:

• 🛒 4 e-commerce AI business ideas without your own product

• 🪄 Onepage AI 2.0: vibe coding meets no-code website building

• 🐡 Fugu: is this the new AI model approach?

Let's go.

_{Presented by Onepage}

🛒 4 E-Commerce AI Business Ideas Without Your Own Product

We have not shared business ideas in a while. If you have not been around that long: we love looking at the new opportunities AI opens up for normal people.

Many AInauts are always looking for inspiration. This week we stumbled across a post that stuck with us:

Tobi Lütke is the founder of Shopify. Shopify recently introduced an API that gives creative non-techies a way to build interesting commerce experiences. Fine, you still need a bit of technical curiosity. But still.

Behind it is UCP, the Universal Commerce Protocol.

Shopify rolled it out with its new Spring Edition, together with Google, as an open standard.

Translated: AI agents can search the full Shopify catalog with billions of products, build carts, and complete checkout directly.

The real breakthrough is something else. Until now, you needed approval from Shopify for this kind of thing. That gate is gone.

And you do not even need your own shop to build on it. Register a profile, use the public endpoint, done.

We call it the Stripe moment for shopping. Stripe turned payments into a layer everyone could build on. That is now starting to happen with shopping.

To make it concrete, here are a few ideas for using this API and millions of Shopify stores.

4 Ideas You Could Start With

1. The niche buying advisor. Take the sneaker demo and turn it into a business. A chat agent for a tightly defined audience: gear for trail runners, first equipment for new parents, home-office setups for creators, or a gift assistant for people who hate choosing gifts.

   It curates from the huge catalog and guides people directly to purchase. You earn through commission, a small subscription, or as a smart lead magnet for your own business.

2. The honest best-value finder. Build your own niche version of Stiftung Warentest or Consumer Reports in the area you know best. Market it exactly that way. Trust is the lever here.

3. The gift concierge via WhatsApp. Simple flow: connect the API to a WhatsApp or Telegram number and define a bot workflow like: describe the person, budget, occasion, and taboo gifts.

   Then the agent suggests five suitable products. Around Christmas or Mother's Day, that is almost a no-brainer.

4. Shopping directly inside your content. You have an audience, a newsletter, or a community? Build shopping directly into the context. You recommend something, and the reader buys without jumping through five extra steps.

Our Take

Yes, implementation still gets a bit technical. But as we wrote on Monday: building small applications and connecting APIs is no longer nearly as hard as it used to be.

We will build a dummy version of this ourselves.

And yes, users could also do this with ChatGPT and a Shopify integration.

These ideas are more arbitrage than built-to-last businesses.

But with more than ten years in affiliate marketing, we can tell you: these things can be worth it. With AI, it becomes even more relevant who manages to capture attention at the exact moment people want to buy.

If you already have a niche or an audience, the hardest part is done. The rest is now an API. We all become marketers, until maybe even that becomes unnecessary. 😉

Investors see ANOTHER return from Masterworks (!!!!)

That’s 6 sales in 7 months. 29 all time. And the performance?

16.5%, 17.6%, and 17.8%, net annualized returns on sold works held longer than one year (See all 29 at Masterworks.com)

It’s not from stocks, private equity, or real estate… it’s from contemporary and post war art. Crazy, right?

With Masterworks, you don’t need to be a BILLIONAIRE to invest in multi-million dollar art anymore.

Historically, the segment overall has had attractive appreciation and low correlation to stocks.*

Masterworks targets works featuring legends like Banksy, Basquiat, and Picasso, identifying what they believe to have significant long-term appreciation potential, not just at the artist level but at the level of individual artworks.

As one of the largest players in the art market, with $1.3 billion invested over 500 artworks, they pass critical advantages through to their 70,000+ members to add art to their portfolios strategically.

Looking to diversify your investments in 2026?

Click here to skip the waitlist

^{*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.}

🪄 Onepage AI 2.0: Vibe Coding Meets No-Code

Note: Onepage is a German company, and a few screenshots in this section may show German interface labels.

Almost every one of us eventually has to build a website or funnel.

Privately, for a business, or for a side project. Thanks to AI, that no longer requires a technical degree.

You describe what you want and get something good-looking within minutes.

But many AI tools have a problem, especially for non-techies: in the end, they give you code. And as soon as you want to change one small thing, the drama starts.

Prompting tiny text changes often takes longer than just making them yourself.

That is exactly why we showed you Onepage a year ago, a website builder we have used for our own pages for years.

The great part back then: AI builds the page, and you can adjust it afterward without touching code.

That was cool, but it had two small catches:

• Adjustments were still fully manual.

• The Onepage AI was still locked into fairly rigid rules.

Now Onepage AI 2.0 is in beta, and it brings a lot of new capabilities.

We tested it directly and had it build an "AInauten AI Consulting" page from a prompt.

One paragraph about what we wanted, and the AI got going.

It built section after section in AInauten purple: hero, service cards, timeline, and animated counters.

And while it keeps building in the background, you can already edit the finished parts in the preview on the right.

That is the vibe-coding part: the AI writes real code in the background, while you work only with what appears in the visual preview.

Without a lot of prompting, Onepage's marketing know-how gives you genuinely strong websites, especially when conversion and structure matter.

You can immediately change text manually and very quickly. Colors, buttons, and similar details: click in and adjust.

Beyond the completely new experience, the AI 2.0 update brings several useful features. The most impressive one for us is Promptable Controls.

Every generated page contains different sections.

For each section, the AI also builds its own control panel where you can adjust colors and other details.

The kicker: if you want more controls, you can simply ask it to build them.

We asked it to add a panel for accent color, headline, and button text to the hero. Seconds later, we had exactly that: a visual color picker and input fields.

You say once what you want to control, get the sliders and fields for it, and can then visually change the design.

That is the real jump in 2.0. The first AI wave was fast, but afterward you were stuck. Here, AI builds the site and gives you control at the same time.

Vibe coding, meaning you describe what you want, meets no-code, meaning everything stays visually editable. Add real animations and small interactions, not just static blocks.

And both parts work together. You can start from a finished template or prebuilt Onepage section and upgrade it by prompt. Or you can create a section from scratch and still edit it visually afterward.

Onepage is not just a page generator either. What the AI builds lives directly inside Onepage. You can host it, publish it, and continue with CRM, SEO, and the usual operating work.

Sure, betas can wobble, but for us it ran quite stable.

And as always with Onepage: German company, GDPR-compliant, hosted in Germany.

If you need a website or funnel and you are not a techie, take a look at the new Onepage AI 2.0 beta.

P.S. You can even build pages directly from Claude now, because Onepage has added an MCP connector. We will show you how that works in detail soon.

Your best prompts are the ones you'd never bother typing.

The detailed ones. The ones with examples and edge cases. Wispr Flow lets you speak them instead — clean, structured, ready to paste into any AI tool. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

🐡 Fugu: Another Model Team, and Our Take on It

Last week we showed you OpenRouter Fusion. This week, Fugu from Sakana AI is making headlines, and the principle is similar: not one model, but a whole team of models working together.

We include it because there seems to be a trend: more providers are trying to get top performance not from one giant model, but from coordination, routing, and model collaboration.

Does that mean we no longer need the top models from OpenAI and others?

As always, we want to give you a short assessment before you spend money on API models.

After our first tests, honestly: you can skip Fugu for now. Still pretty buggy, and in our tests not better than Opus and similar models.

The approach itself is strong though, both with Fugu and with Fusion. A new way of building AI models is emerging here, and that remains interesting.

If you still want to play with it: Sakana is once again not available in the EU, but you can test Fugu through OpenRouter.

For everyone who enjoys trying new models: take a quick look, but keep expectations under control.

That is it for today. We will be back next week with new, useful things to try.

Until then!

Reto & Fabian from AInauten

Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

You know those small subscriptions that quietly charge every month because they solve one tiny job well enough? We know that feeling too. Ouch.

That is exactly where AI is getting interesting. More and more often, you can have small tools built for your own workflow instead of renting yet another SaaS. Today we are looking at those mini-apps, ChatGPT as a real workspace, and Germany's AI location index.

Here is what we have for you today:

• 👨‍🚀 SaaS Killer: Your Own App Instead of Another Subscription

• 🤖 ChatGPT Is Moving from Answer Box to Workspace

• 🇩🇪 Germany's AI Location Index 2026

Let's go.

Wake Up Smarter About AI.

The president of OpenAI. The CEO of Google. The founder of LinkedIn. We talk to the people building AI. Every morning, we put what they told us in your inbox.

What they're actually worried about. What they're betting on. What's coming next. Delivered in five minutes, before your first meeting.

With one email, and only five minutes, you can stay ahead of 99% of the world.

Join 75K Readers →

👨‍🚀 SaaS Killer: Your Own App Instead of Another Subscription

Ralf killed a SaaS inside our AI Automation Community. A completely normal work moment for him, which is exactly why we like the example so much.

For meeting notes, he could have paid for another subscription tool. Instead, he built and shared HyperScribe: a small desktop app for macOS and Windows that uses Deepgram to transcribe audio, separate speakers, and do the job for roughly 30 cents per hour.

The next day, he followed up with his "Confessions of a SaaS hitman." With HyperRec, he built a small tool that records microphone and speaker audio directly. Another subscription replaced by one focused workflow.

The motto: tool annoys you, subscription annoys you, agent builds a better shortcut. Done.

Can you do this too?

We say this so often now that we are starting to annoy ourselves. Sorry. But the point matters.

You do not need to know how to code. The technical part is secondary. You need a job that you can describe.

The real blocker is rarely the tech. It is the moment where you dare to give your agent a real job.

And if the agent claims it cannot do it, get a bit stubborn: "I know you can do this. I believe in you."

Not a joke. Sometimes that kind of push is exactly what helps.

Anyone can prompt an app

Second example before we build one ourselves: last week I visited an old friend from my teenage years. Martin works as a building services planner and engineer. IT, programming, and vibe coding are not exactly his natural habitat.

Our starter prompt: "I am here with Martin. He works as a building services planner and engineer. I want to show him what is possible with Codex. Give me 10 ideas for a small mini-app that would be relevant to his work."

A few minutes later we had the idea: a pump pre-selection tool that suggests suitable products from relevant manufacturers based on project constraints.

A bit of prompt ping-pong later, the first version worked. Then we asked:

Great, this works.

How could we make the app 10x better now? What features would help? What would you optimize?

Analyze everything in detail.

Give me 10 to 20 suggestions for useful improvements and prioritize them.

More than half of the suggestions were implemented while we sat in the garden talking about old times and future plans.

With a few clicks, we deployed the app for free on Cloudflare Pages, password-protected and ready.

On the way home, his WhatsApp arrived. In plain English: I can actually use this. I got a taste for it. That is exactly what you want from a first prototype.

Your SaaS-killer prompt for Codex App, Claude Code, or Cowork

If you want to start, start small. Our first "app" was a simple browser extension, painfully prompted with ChatGPT back then. Since then we have built dozens of small and large tools with AI.

You still do not need to have built an app before. Install the Codex App from OpenAI or the Claude Code / Claude Cowork desktop app as your base.

1. Start a new chat with an empty folder for the project.

2. Use the best model with a high thinking budget and begin in planning mode.

3. Paste this prompt, or use the detailed version here, and get started.

I want to build a small local tool that saves me a subscription or a paid workflow.

Find at least 5 possible candidates from my workday. Rate them by savings potential, frequency, technical complexity, and usefulness.

Then choose the best candidate and write a short PLAN.md and TASKS.md for the build.

Important: do not write code yet. First understand the job, the input, the desired output, and the acceptance criteria.

Tip: you can also ask ChatGPT or Claude first, because that chatbot probably knows you better than a freshly installed Codex or Claude app. Then add: "IMPORTANT: write this so that I can paste it directly into Codex / Claude Code."

The most important sentence is not at the top. It is further down:

Do not write code yet.

Unsexy, but decisive. Many agents sprint off, build half an app, solve three side problems, and forget why you started.

This prompt forces the agent to understand your situation first and save the context in files. PLAN.md and TASKS.md are not bureaucracy. They keep the agent oriented after 40 minutes.

Anthropic recommends exactly this kind of clean task split in its agent guides: clarify goal and acceptance first, then automate. In practice, that is the difference between "looks cool" and "still works tomorrow."

While testing the prompt, we built a small app that turns a webinar transcript into working materials for us: summary, infographic prompt, mailing idea, and course notes.

Our take: an app is not a SaaS, but it is a first win

If you have a button after two hours that takes a real file and produces a useful output, you have learned more than from ten tool demos.

Maybe you cancel the subscription afterward. Maybe you keep paying for it, but finally for a good reason.

Real apps need more later: security, data, tests, hosting, support, payment, and the boring parts that keep things alive.

For today, aim smaller: build one tool that solves one real problem in your day. Then take the win.

You built something while others were still collecting examples.

P.S. If you need more inspiration, this X post includes twelve prompts where the author replaced smaller AI tools, from grammar checks to meeting summaries and SEO briefings.

Wake Up Smarter About AI.

Most AI news is a waste of time. The Future Today is a daily 5-minute read focused on what actually matters.

You’ll get exclusive interviews with the CEOs, researchers, and builders shaping AI. Plus top stories broken down simply and practical advice you can immediately apply.

Read the newsletter trusted by teams at NVIDIA, Google, Anthropic, Meta, Dell, and Salesforce.

Join Free Today →

🤖 ChatGPT Is Moving from Answer Box to Workspace

OpenAI is pushing ChatGPT and Codex in the same direction: away from the pure answer window and closer to your actual work.

The effect may matter more than the next model rumor: you have to trigger fewer routines manually.

ChatGPT agents move into the shop window

On the new Scheduled page in ChatGPT, you can see scheduled tasks, check the next run, pause, resume, edit, or delete tasks.

Twitter tweet

Tasks can handle reminders, recurring work, and monitoring. Less "ask me again later." More "I will tell you when something happens."

Daily briefings, email monitoring, concert alerts, sale alerts: ChatGPT can watch the things you used to check manually.

If you are not using Scheduled Tasks yet, click here and create one. Takes five seconds.

OpenAI even removed the popular Premium feature Pulse. We read that as a clear signal: the chat should stop only answering and start taking on small obligations.

The smaller ChatGPT updates point in the same direction and make the chat window feel more like a work environment.

OpenAI's agents want to get closer to your computer. Let them.

The new Tasks page is currently meant for ChatGPT on web and mobile, not the desktop app or Codex. That detail matters, because OpenAI is also moving Codex closer to the computer. See our earlier post. The relevant pieces are:

• Computer Use for controlling macOS, Windows, and apps

• Codex Chrome Extension for logged-in browser contexts

• Codex's own browser based on OpenAI's Atlas browser

• Memories and Chronicle as optional context layers

Many productive AI workflows are not blocked because the model is too dumb. They are blocked because the work lives in other tools and surfaces: Gmail, LinkedIn, Salesforce, WordPress, internal admin tools, local apps.

More Codex features have started rolling out for us over the past few days. That is where things get interesting. And risky.

If Codex can use Chrome, it can see logged-in pages. If Codex uses Computer Use, it can click, type, and operate windows. Powerful, yes. Also not a casual prompt.

Record & Replay is the next big lever, just not everywhere yet

Record & Replay is next! The idea is brilliant: you show Codex a recurring workflow once, and Codex turns it into a reusable skill.

Twitter tweet

Still, the direction is clear: from "describe the workflow" to "show me the workflow once." That is much closer to how normal people work.

ChatGPT's market share falls below 50% for the first time

Outside pressure explains why OpenAI wants to consolidate the interface. ChatGPT is still huge, but Sensor Tower data says its former 90 percent share has fallen to 46.4 percent of the global AI-assistant market.

Gemini sits at 27.7 percent, Claude at 10.3 percent. ChatGPT is still enormous with more than 1.1 billion monthly users. The market is not collapsing; it is growing and spreading out.

Gemini is attached to Android and Workspace. Claude keeps growing with power users. Microsoft is rolling out Copilot Cowork globally for Microsoft 365 Copilot customers.

Agents are becoming small workflows with a budget line.

Our take: delegate your tasks to agents

For your daily work, the decisive question is not whether ChatGPT, Claude, Gemini, or Copilot is ahead in a benchmark. The question is: which agent can you give a job to without feeling nervous afterward?

Waiting is not our favorite strategy. Start with these features:

• Scheduled Tasks are the easiest entry point: small routines that do not need a fresh prompt every time.

• Codex Computer Use is the next stage: tasks inside real interfaces.

• Local mini-tool: one small workflow that replaces a subscription or recurring manual task.

Have fun with it. Ideally, it gives you a real productivity bump.

🇩🇪 Germany's AI Location Index 2026

Berlin is Germany's strongest AI location, at least according to the new KI-Standort-Index Deutschland 2026.

Berlin lands in first place with 84.7 points, ahead of Munich at 77.9 and Stuttgart at 66.1. After that, the gap gets steep: Hamburg reaches fourth place with only 43.3 points.

The ranking compares 37 cities across four pillars:

• open AI jobs

• research and education

• startup ecosystem

• salary after rent

So far, so measurable.

But what is an AI location, really?

A few lighthouse projects are not enough. A city becomes relevant when several things happen at once: universities train AI people, companies hire them, startups build products, and people can afford to live there.

We think the ranking is a useful reality check. Berlin, Munich, and Stuttgart are not at the top because they make the loudest AI promises. They are visible because more of the system already exists there: jobs, research, startups, and economic pull.

The German question becomes European very quickly. Even the best location remains dependent when models, chips, cloud access, and platform rules are controlled elsewhere.

The bigger picture is this DW video: US limits AI access - What it means for Europe. It asks whether AI access itself is becoming a geopolitical tool.

P.S. If you want to hear how one of the most important American AI CEOs currently sees the world, the long interview with Anthropic CEO Dario Amodei is a good bonus: here you go.

Done. Today's issue had a lot of agent future, and a lot of agent present, but the practical assignment stays grounded: pick one recurring job you can describe cleanly. Let an agent plan first. Then let it build.

And if one subscription disappears along the way: even better.

See you in the next issue.

Reto & Fabian from AInauten

AIHOY and happy weekend, AInauts!

Maybe you missed a few AI news, tools, and practical ideas this week. Or maybe you have just joined us. No problem.

Here are our highlights from the past week:

• 👷‍♂️ How to Find Real Jobs for Your Agents

• 🧨 Fable 5 is gone: why local models suddenly matter

• 👌 OpenRouter Fusion: Fable-level output from several models?

• 🔓 The 3 Stages of AI Independence

• 🛠️ OpenWork: the open-source alternative to Cowork

• 🎯 Meta-prompting: let the AI write the assignment itself

And at the end, you get the most important Quick News, so you can catch the relevant developments in one pass. Ready? Let's go.

👷‍♂️ How to Find Real Jobs for Your Agents

The technical setup is rarely the bottleneck anymore. Creating skills, testing MCPs, starting a browser agent: all of that works surprisingly well now. The real bend in the road comes afterward, with one plain question: what should the agent actually do for you each morning?

In our AI Employee Bootcamp, we see exactly this gap. That is why we give you three prompts: find the best agent jobs from your real work, turn them into a clean briefing, and set up the agent directly on your machine. Plus we clear up the endless confusion around skills, MCPs, and agents.

Our advice: build one boring agent this weekend. One you can honestly evaluate after seven days. Did it save you 20 minutes every morning? That is enough for a start.

👉 Go to the three agent prompts »

🧨 Fable 5 is gone: why local models suddenly matter

For three days, Claude Fable 5 was the model that made everything else look old: huge context, long autonomous tasks, finally a model that did not need hand-holding after every second sentence. Then the U.S. government pulled the plug.

On June 12, the Trump administration demanded that access to Fable 5 and Mythos 5 be blocked for all foreign nationals, even Anthropic employees in the U.S. There was no clean way to enforce that, so Anthropic switched both models off for everyone. The trigger was reportedly a jailbreak.

The lesson for us is uncomfortable: the strongest tool in your stack can disappear tomorrow, and you cannot stop it. A kill switch you do not control. That is exactly why local and open models are no longer hobby gear. They are basic infrastructure.

👉 Read the full Fable story »

👌 OpenRouter Fusion: Fable-level output from several models?

While everyone is mourning Fable, OpenRouter is making a cheeky argument: maybe one single model does not have to do everything anymore.

The new OpenRouter Fusion is basically a small model panel. You ask a complex question, several models work on it in parallel, a judge model compares the answers, and the system builds the final result. That is closer to good teamwork than classic routing.

For short prompts, Fusion is overkill. But anywhere a single answer is dangerously convenient, such as research, strategy, or comparisons, it gets interesting.

👉 Read our Fusion piece »

🔓 The 3 Stages of AI Independence

After the Fable shock, the same question kept coming up: how do I make myself less dependent? We sorted our answer into three stages.

Stage 1 is familiar from us: Files over Tools. Your knowledge lives in folders, not locked inside one app, so you can move from Claude to Codex to Antigravity without losing your work. Stage 2 makes you independent from one subscription: an OpenRouter key gives you access to hundreds of models from more than 75 providers. Stage 3 is the radical version: local models that run on your own machine.

Independence does not mean replacing everything with free tools. It means never being tied to a single provider again. Usually you give up a bit of frontier-model performance. For many use cases, that trade-off is completely manageable.

👉 Go to the 3 stages »

🛠️ OpenWork: the open-source alternative to Cowork

Chatting with a local model is nice. But many of us now work differently: the AI reads and writes in our folders and works through tasks step by step. Cowork-style work, basically.

That now exists in an open version too. OpenWork describes itself as an open-source alternative to Claude Cowork, a desktop app for Mac, Windows, and Linux. It connects to your local folders, handles multi-step tasks, supports skills, templates, and permissions, and works with your own OpenRouter key.

We tested it quite a bit and find it very interesting. Your folder stays the source of truth, and you pick the model from a dropdown. That way of working belongs to you. No subscription can cancel it for you.

👉 Go to OpenWork »

🎯 Meta-prompting: let the AI write the assignment itself

To close, here is a small hack that connects directly to last week's loops and prompt tips. Meta-prompting used to be a big topic in our trainings: prompts that write better prompts for us.

Now we can lift that idea into the agent world. Instead of typing a clean briefing yourself, you let the AI write its own /goal. You give it the task, context, and boundaries, and it turns that into a main goal, success criteria, and stop rules. You only approve it.

That means you delegate the work and the briefing. And honestly, the AI often knows its own mechanics better than we do. This gets every task started cleanly, without you writing a full brief first.

👉 Go to the meta-prompting template »

AI News Quickie: the HAI-lights from the industry

AI never sleeps. This week was all about record deals and the biggest IPO ever, agents that suddenly pay and work on their own, and governments trying to slow AI down. Here are the most important stories.

🏛️ Anthropic, Claude, and the Fable aftermath

• Fable 5 is still offline, but Anthropic sounds confident it will return soon. The White House reportedly feared Chinese access and gave Anthropic 90 minutes to shut it down.

• More than 80 security leaders from Adobe, Zoom, Nvidia, and others are asking in an open letter for the restriction to be lifted. Some had warned in April how dangerous Mythos could be.

• Anthropic is opening an office in Seoul and has signed a memorandum of understanding with Korea's science ministry.

• Through an alliance with Tata Consulting Services, Anthropic is putting Claude in the hands of 50,000 consultants and moving into regulated industries. In other words: where the money is.

OpenAI under constant fire

• OpenAI is making Shopping Ads in ChatGPT almost automatic: merchants connect their catalogs, and the system builds the ads.

• Smart move: with Deployment Simulation, OpenAI replays old conversations through a new model before each release to predict bad behavior.

• With LifeSciBench, OpenAI tests how well AI can handle real life-science research: conflicting data, incomplete evidence, and experiment design.

• Companies now get Spend Controls in ChatGPT Enterprise: credit limits per team and person. Usage finally becomes visible.

• OpenAI is buying cloud company Ona, formerly Gitpod, so Codex can run long coding jobs in the cloud even when your laptop is closed.

• After the confidential S-1, OpenAI is now reportedly heading toward a Q4 IPO with Goldman Sachs and Morgan Stanley.

• GPT-5.2 story: all chats are automatically moving to GPT-5.5.

🖥️ Google between launches and cleanup

• With the Open Knowledge Format, Google is making the LLM wiki pattern official: knowledge as Markdown instead of proprietary catalogs. Files over Tools says hello.

• The Antigravity CLI now replaces the old Gemini CLI. For many teams, that is a forced migration that breaks existing automations first.

• DeepMind will treat its own agents like potential insider threats: rights are granted step by step, and monitoring flags suspicious behavior. Trust is good, sandboxing is better.

• The June Pixel Drop brings new features such as text-to-video with Gemini Omni, Screen Reactions, and live speech translation to phones. AI keeps moving into the pocket.

• Google is suing a Chinese AI fraud network that used Gemini to build scam pages and defraud hundreds of thousands of victims. AI helps both the attack and the defense.

• Google's AI Overviews are telling users that horror-fiction creatures are real.

Agents learn to pay, work, and remember

• Coinbase and AWS now let agents pay with crypto: through x402, a CloudFront provider can invoice the AI directly.

• Visa follows and lets agents in ChatGPT handle real purchases, including spending limits and tokenization. Two payment rails, one goal: your agent pulls out the card.

• Microsoft Copilot Cowork is live globally, and it runs on Claude Opus 4.8.

• OpenAI is launching a Partner Network with 150 million dollars and Accenture, BCG, and McKinsey. 300,000 consultants are supposed to carry AI into companies.

• Bedrock gets AgentCore Web Search and a managed knowledge base. AWS is building the workbench for enterprise agents.

• Perplexity gives its agent a learning memory: Brain builds a context graph overnight from completed work and remembers what worked.

• Sakana AI introduces Marlin, a virtual CSO: an agent that researches for up to eight hours and produces 100-page reports plus slides.

The deals are getting wilder

• SpaceX, now merged with xAI, is swallowing coding star Cursor for 60 billion dollars in stock. Elon is filling the gaps in his empire.

• Just days earlier, SpaceX itself went public: the largest IPO ever raised 75 billion dollars, and the stock jumped 19 percent on day one.

• Salesforce is buying support platform Fin, formerly Intercom, for 3.6 billion dollars.

• Snap is spinning out its expensive gen-AI video team into a separate company called Dotmo. Too expensive in-house, too valuable to throw away.

• Mistral, Europe's answer to U.S. dominance, is negotiating 3 billion euros in fresh capital and aiming for a 20 billion valuation.

• Perplexity is raising 200 million dollars to keep scaling its AI browser Comet.

🌍 Power, money, and infrastructure

• Where is Germany's AI capital? The new AI Location Index 2026 compares 37 cities: Berlin leads clearly with 84.7 points ahead of Munich. Freiburg has the highest job density, and AI Engineer is the most sought-after profile.

• At the G7 summit in Evian, Altman, Amodei, Hassabis, and others sat together in front of heads of state for the first time. Topic: AI sovereignty and fear of the kill switch. After the Fable drama, no surprise.

• China is planning a 295 billion dollar national AI data-center grid, 80 percent powered by domestic chips. Nvidia is deliberately kept outside.

• Oracle reports an AI order backlog of 638 billion dollars, up 363 percent. The stock still fell: that much backlog has to be financed and built first.

• GitHub is routing traffic to AWS because AI coding agents are overrunning the infrastructure: 275 million commits per week and availability temporarily below 89 percent.

🛡️ Politics, society, and law

• The White House is making its AI rules in real time. The Fable shutdown was the preview, and a clear line is still missing.

• Amazon employees say they are being pushed out because they supported data-center limits.

• A new Pew study shows that two thirds of Americans think AI is moving too fast, and 40 percent expect more harm than benefit. Even younger people are skeptical while still using chatbots heavily.

• Actors' union SAG-AFTRA has ratified a four-year contract with the studios, including strict rules against digital doubles. More than 90 percent voted yes.

• In the RIAA case against Suno, discovery reportedly shows millions of copied songs. A fair-use decision is expected in July.

• Bitter punchline: the world's leading deepfake expert no longer trusts his own eyes. Media literacy is becoming a survival skill.

🎬 Tools, hardware, and the creative world

• xAI's Grok Imagine Video 1.5 jumps to the top of the video arena, with sound and at a fraction of Sora's price. The AI video price war is on.

• Adobe is putting its Firefly assistant into Premiere, Illustrator, and InDesign too: sort clips, clean up layers, and build brand kits by instruction.

• Meta is rebuilding Facebook around AI Mode: instead of scrolling, you ask in normal language and get an answer from public posts, groups, and Reels.

• Snap is selling its AR glasses Specs to everyone for a hefty 2,195 dollars.

• Xreal joins in too: the Aura glasses run on Android XR with deep Gemini integration and weigh under 95 grams.

• Epic is building generative AI directly into Unreal Engine, with Claude, Gemini, and Codex via MCP.

• Pinterest launches the experimental shopping app Ask Pinterest plus AI ad tools and its own MCP. Inspiration turns into agentic shopping.

• At Cannes Lions, NVIDIA and partners are reshaping advertising and marketing with AI.

Done. But no need to be sad. The AInauten will be back soon with fresh fuel for you.

See you Monday with a new round of news, hacks, and insights.

Your AInauten, Fabian & Reto

P.S. Follow us on the socials: X, LinkedIn, Facebook, Instagram, YouTube, TikTok.

Hi AInauts,

Welcome to the latest issue of your favorite newsletter.

We could spend today talking about how much richer Elon Musk just got and how the AI company Cursor was reportedly bought for 60 billion dollars. But that is not really the point today.

Tiny fun fact anyway: technically, your net worth is closer to Bill Gates' than Bill Gates' is to Elon Musk's.

But the bigger story is still the U.S.-government-triggered shutdown of Anthropic's Fable 5 model.

After Monday's issue, a lot of questions and follow-up topics came in.

So today we want to pull the whole thing together under one theme: independence.

Because this is not just an Anthropic story. It is a broader issue that will hit all of us in some way. As always, we are looking at it as practically as possible.

Here is what we have for you:

• 🔓 Fable 5 is gone: the 3 stages of AI independence

• 🛠️ OpenWork: the open-source alternative to Claude Cowork and Co.

• The new meta-prompting and how to use it

Let's go.

The ones showing up in LLMs convert 3× better than Google

They optimized for LLMs, not just Google.

FAQs. Comparison pages. Transparent pricing. LinkedIn presence. These aren't vanity plays. They're what gets you cited in ChatGPT, Gemini, and Claude when your buyers are researching, your investors are looking, and your future hires are deciding where to work.

Download the free AEO Playbook for Startups from HubSpot and get the exact checklist. Five minutes to read.

Unlock the playbook

🔓 Fable 5 is gone: the 3 stages of AI independence

The whole Fable 5 story is genuinely fascinating.

Here is one possible view through the lens of the U.S. government, via David Sacks.

This post showed us just how meta the whole topic has become:

The moral of this fable is pretty clear to us: set yourself up to be as independent as possible in AI.

That is why we pulled together our different approaches again, in a more concrete way.

We see independence in three stages.

Stage 1: independent between the labs

You already know stage 1 from us: Files over Tools.

Your knowledge lives in folders, not locked inside one app. That means you can move from Claude to Codex to Antigravity without losing the work.

Short version: work with AI in local folder structures. Let it document progress there. Keep projects in project folders. Out of endless chats, into local Markdown files.

Make sure your knowledge stays with you, not inside someone else's tool.

But what happens when the problem is not one annoying lab, but rising subscription prices, canceled access, or blocked models?

Stage 2: independent from one subscription and one model

This is where OpenRouter comes in.

In three sentences: one account, one key, access to hundreds of models from more than 75 providers, both closed and open. Claude, GPT, Gemini, MiniMax, all in the same dropdown.

You top up credit instead of committing to another subscription.

The point: you need fewer subscriptions and become less dependent. For almost every model, there is now an almost-as-good alternative.

We already wrote about how OpenRouter's new Fusion model gets close to Fable 5-level performance. Use that for your more complex work.

OpenRouter also has a leaderboard. The current favorite as a daily driver and cheaper alternative to frontier models is MiniMax M3.

Another important piece of independence at this stage is dependence on the providers' own apps.

Claude Desktop with Claude Code/Cowork or OpenAI's Codex are seriously powerful. But without an account there, you also lose access to the app experience.

But we have a very interesting alternative below.

Stage 3: independent from the cloud

The final stage is the most radical: models that run entirely on your own computer. Offline, private, free.

We have already written about how to check what your computer can actually handle.

Here is the easiest app for getting started with local models: LM Studio.

One app, no Docker fiddling. You download a model and chat locally. With an official plugin, you can also connect your OpenRouter key in the same window.

That gives you local and frontier models side by side.

If your machine can handle it, we recommend Google's Gemma models.

But we should be honest: local models are not at Fable level. The really good ones still need a powerful machine, and they are still quite far away from the top models.

For private, sensitive, simple, offline work, though, they are unbeatable.

Our take: build your sovereign AI stack

Independence does not mean replacing everything with free tools.

It means never again depending on a single provider.

With Files over Tools, OpenRouter, LM Studio, and the right model choices, that is absolutely possible.

You usually have to dial down your expectations compared with the top models, but for many tasks that trade-off is perfectly manageable.

A small quick hack to close this section

This week we got a question: what do I do with my old Fable 5 chats that I can no longer continue?

Here is a simple workaround:

• Open a new chat.

• Ask Claude whether it can read the old Fable 5 chat named [INSERT NAME].

You find the name at the top of the conversation.

Claude can then pull the information from the old chat and you can keep working with another model.

AI is in 60% of engineering work, but only 20% can be handed off without someone babysitting the output. Join this live webinar on June 24 (free) to see how top teams are using a context engine to level up.

Register Now

🛠️ OpenWork: the alternative to Claude Cowork and Co.

Above, we already teased why it matters to become less dependent on apps like Claude Cowork and Co.

The tool we want to show you today deserves its own section.

Chatting through LM Studio is one thing. But many of us now work differently. The AI reads and writes in our folders and completes tasks step by step.

Cowork-style work, basically.

That is exactly why this open-source alternative is interesting.

What OpenWork is

OpenWork describes itself as an "open-source alternative to Claude Cowork." It is a desktop app for Mac, Windows, and Linux.

It connects to your local folders, reads and writes files, handles multi-step tasks, and supports skills, templates, and permissions. In other words: the Cowork logic, but through an open app.

Of course, you can also connect MCPs and similar tools.

From a usability perspective, it feels pretty close to the other desktop agent apps.

That is why we keep saying in the AI Employee Bootcamp: learn one app and this way of working, and you can transfer the pattern to all of them.

Why it matters for independence

OpenWork runs with your own OpenRouter key. That is stage 2 from above. And it can even connect to local models through Ollama, which is stage 3.

Your folder stays the source of truth. You choose the model from the dropdown.

Plain English: this way of working then belongs to you. No subscription can take it away from you.

You can download the desktop app directly from OpenWork after creating a free account.

Our take: a solid alternative for Cowork-style work

After some testing, we genuinely like OpenWork.

It is a very good alternative and definitely makes you more independent.

It is really not rocket science. Give it a try.

The new meta-prompting: let the AI write the assignment itself

To close, here is something that connects directly to last week's loop topic. A small prompt hack for today.

If you have followed us for a while, you may remember that meta-prompting used to be a big topic in our trainings.

Basically, we used prompts that wrote better prompts for our actual tasks.

Today's hack adapts that idea to the agent world.

Twitter tweet

First, use the /goal skill in Claude or Codex so the AI writes a clean /goal for itself. You then approve it, and off it goes.

That is meta-prompting for everyone. You are not only delegating the work, you are also delegating the briefing. And the AI often knows its own mechanics better than you do.

Here is a prompt template you can copy directly:

Before you begin, write your own /goal for this task.

Task: [describe the task]
Context: [files, docs, requirements, or links here]
Boundaries: [scope, style rules, deadlines, what to avoid]
Done means: [what the result should look like]

Return:
1. Your main /goal.
2. 3 to 5 success criteria.
3. Boundaries that should not be crossed.
4. If you start helper agents, write a separate /goal for each one.
5. Ask me whether I approve or want to adjust the goals.

Do not start until I approve.

Then you only approve or briefly adjust the goals. And every task starts with a clean briefing you did not even have to write yourself.

That is it for today. A pretty independent issue, all things considered.

Thanks for reading us, independent AInauts.

See you next week.

Reto & Fabian from AInauten

Hello AInauts,

Welcome to the latest issue of your favorite newsletter!

Today we are not starting with the next tool. We are starting with the question: What specific task should your agent handle for you every morning? With our prompts, you will find your first practical agent in three steps and can test it directly in your everyday work.

After that, we look at Claude Fable 5, which briefly showed what the AI future can feel like before it disappeared overnight. At the end, we test the counter-idea: if the supermodel suddenly vanishes, can a model team step in?

Here is what we have packed for you today:

• 👷‍♂️ How to find real jobs for your agents

• 🚨 Claude Fable 5 shutdown makes local models more important

• 👌 OpenRouter Fusion: Fable-level output from multiple models? Almost...

Let's go!

Learn AI in 5 minutes a day

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

👷‍♂️ How to find real jobs for your agents

In recent AInauten sessions we are seeing an interesting pattern: the technology is often no longer the bottleneck. Creating skills, connecting files, testing MCP servers, starting browser agents... all of that now works surprisingly well.

The break comes after that: what is the agent actually supposed to do for you?

An agent becomes useful once it can handle a recurring task with enough context and clear boundaries, reliably.

So today we want to look at how you can find concrete candidates from your real work or everyday life. For most people, this is where it gets interesting.

Step 1: Prompt "Find my best agent jobs"

Let's get straight into it, without a long preface: copy this prompt into the chatbot you use most.

Analyze my goals, working style, recurring tasks, and tools based on everything you know about me.
If you are not sure, ask me concrete multiple-choice questions.

Then find at least 10 concrete tasks that I could delegate to an AI agent, workflow, or automation.

Rate each candidate by:
- repetition
- time saved
- data access
- risk
- measurability

Choose the 3 best starting candidates.

For each candidate, describe:
- desired outcome
- required context
- allowed actions
- stop rules
- human review point
- success criterion after 7 days

Rules:
No vague ideas. Prefer recurring screen work.
Sending, deleting, paying, and customer contact stay with the human.

Step 2: Prompt "Define the agent job"

Many people have never actively built an agent. Totally fine. It sounds complicated. But because it does not have to be complicated, we will do it together today.

Now choose one agent job from the result of step 1.

Do not pick the biggest, coolest, or technically wildest suggestion. Instead, choose one or more suggestions where you immediately know what a good result would look like.

Good agent jobs often sound simple:

• Prepare draft replies for new support tickets in your style every morning

• Reconcile open invoices in a Google Sheets list every Friday

• Collect questions and feedback from chat and transcript after a webinar

• Prepare five social media posts from the bookmarks of the last few days

That is screen work with context. And that is exactly where the first test is worth it.

Use this prompt after you have selected the first candidate:

Take the best [or selected] candidate from the list.
Turn it into a clean briefing (prompt) for a desktop-agent test.

Work like this:
1. Describe a useful test run and the context required for it.
2. List which local files, folders, websites, or tools the agent needs.
3. Write the concrete task for a desktop agent such as Codex or Claude Code/Cowork.
4. Define stop rules: what may be done, and what may not?
5. Define the review result that shows whether the test worked.
6. Create a checklist to repeat the run cleanly.

Rules:
- Read and plan first.
- No action without approval. No deleting, no sending, no paying.
- Prefer a small test with evidence over a big automation without control.
- If something is unclear, ask instead of guessing.
- Explain all steps simply and traceably.

Step 3: Set up your first agent

Now it gets practical: to set up the agent, you need to install a desktop-agent app such as Codex by OpenAI (usable even with a free account) or Claude Code/Cowork (from the Pro account upward).

Once you have done that, you can in the desktop app start a new chat and paste in the briefing/prompt from step 2 to begin the agent setup process.

At its core, you will be asked to answer questions, grant access, and confirm actions so the agent can carry out the task you gave it.

You will probably also run into terms such as skills, tools, MCP, and so on. We will sort out what those mean in a moment.

Skill, tool, MCP: sorted cleanly once

The desktop agent needs access, but access alone does not create a clean workflow.

Tool, skill, MCP, and context can all show up in the same setup menu. That can create misunderstandings.

Rule of thumb: tools, MCPs, and connectors say what the agent may access. Skill and context say how it should work there.

• A tool grants access. Read Gmail. Check CRM fields. Open a website.

• A skill defines behavior. "Write an email" becomes: "Write an AInauten support reply so the customer understands what happens next."

• MCPs and connectors are the sockets in between. They connect your agent to Drive, Slack, HubSpot, a database, or the browser.

Quick note here: if you install third-party skills, have them reviewed first with NVIDIA's SkillSpector or another tool so you do not install something questionable from a security perspective.

Our take: agents need clean instructions first

If you want to delegate AI work, you have to become uncomfortably precise. Only when the handoff is painfully clear do agents become useful.

So build one boring agent this week. One you can honestly evaluate after seven days: did it save me 20 minutes every morning? That is all it needs at the beginning.

No cable. No subscription. Every match this summer, free.

104 matches. 48 teams. 39 days of football. Right now, streaming means logging into a cable account you don't have or paying for a subscription you'll cancel in August.

Norton Neo is a free browser with a free built-in VPN. No sign-up, no credit card, no catch. Private by default, backed by Norton security. Anti-fingerprinting and ad blocking run quietly in the background while you watch.

Download in 45 seconds. Watch every match for free.

Fast. Safe. Intelligent. That's Neo.

Try Norton Neo Now

🚨 Claude Fable 5 shutdown makes local models more important

For three days, Fable was the model that made us want to push every other model into the corner. Now it is gone, at least for now.

Fable is a preview of the future

Language note: the linked AInauten background article in this section is in German.

The appeal of Claude Fable 5 was that Anthropic made the new Mythos class broadly available for the first time: 1-million-token context, long autonomous tasks, better self-checking, more patience with coding and knowledge work. Exactly the kinds of improvements where many agents have stumbled so far.

Fable felt less like a chatbot and more like a colleague with a lot of coffee and even more IQ. For us, it was a preview of what AI feels like when it does not need to be led by the hand after every second sentence.

And then: the U.S. government pulled the plug.

On June 12, the Trump administration demanded that Fable 5 and Mythos 5 be suspended for all foreign nationals, even foreign Anthropic employees inside the U.S. Because that was practically impossible to implement cleanly, Anthropic had to shut down Fable and Mythos for all customers. WTF?!

The trigger was reportedly a jailbreak that Amazon reportedly confirmed after a White House request . The issue was said to be a simple vulnerability that can also be exploited in other publicly available models such as GPT 5.5.

Because Amazon is also a major investor in Anthropic, we do not quite understand what is going on. Maybe it also has something to do with the rumors that Chinese actors had access to Mythos - who knows?

In any case, Anthropic apologized for the "disruption" and believes it is a misunderstanding. Today there is a discussion. And maybe it is already running again by the time you read this.

Still: a precedent with an important warning.

Security is the most important topic, especially at the breakneck speed things are moving right now (see our post here).

Anthropic had positioned the Mythos class for months as dangerous, powerful, and security-critical, which justified guardrails, trusted access, and high prices.

Some critics say you should not keep telling everyone that your own AI is a biological-weapons and cybercrime factory if you do not want a government to eventually listen seriously.

There is a grain of truth in that...

But if a narrow jailbreak allegation that cannot be publicly verified is enough to rip a production model globally out of workflows, we all learn a hard lesson: the strongest tool in your stack can disappear tomorrow without you being able to do anything. A kill switch you do not control.

via Reddit

Imitating Fable-level behavior becomes a sport

The mood online quickly flipped from hype and cool use cases to dependency shock. The first reaction is the tinkerer path: take available models, build better harnesses, secure prompts, write loops more cleanly, and partially imitate Fable behavior (see our next piece below).

Jailbreaker Pliny had already broken into the model and extracted the system prompt, which was then used in tinkering mode to rebuild Fable-like behavior with the Opus 4.8 Claude model elsewhere.

Sure, a system prompt does not bring the Fable model back, but the idea is not bad.

Sovereign infrastructure: you have a Plan B that actually runs

If you want to play it safe, you need more than workarounds and hope. And if you want control, you need to be able to run local models - or at least use open competing models such as Kimi 2.7, GLM 5.2, and others.

Sovereign AI infrastructure starts small. A local model for confidential standard tasks. A cloud top model for the heavy work. A simple rule for which task may go where. A folder where context, prompts, and checklists live. A test of whether your most important workflow continues if your favorite model disappears.

Tools like llmfit become more interesting in this moment because they ask the right question: What can run sensibly on your hardware?

The tool detects how much RAM, CPU, and GPU your computer has and rates models by quality, speed, fit, and context. It supports providers such as Ollama, LM Studio and others.

Curious what would run on your machine? See the agent section above and simply tell the Claude or Codex desktop app:

Install https://github.com/AlexsJones/llmfit locally.
Then start it for me.

In the Open-Source Arena.ai leaderboard you will find the open cloud models from Kimi, Qwen, and DeepSeek. They rank high on the list and close to the closed cloud providers, close enough that you should seriously test them.

These tests matter, because a benchmark alone says very little about how well something works for your specific use case.

On the free arena.ai platform, not all frontier models are available. If you want to choose from all models for your tests, you can alternatively use OpenRouter (credit system).

arena.ai for model comparison (free)

openrouter.com for model comparison (all models, credit system)

Our take: Fable was a gift, the cutoff is the lesson

We see the shutdown as a dangerous precedent. Safety needs clear processes. Otherwise, caution (or some other motivation) turns into arbitrariness.

The current Fable drama shows that these rules are still missing. Until they exist, total dependence on closed frontier models is an operational risk.

Of course, we do not want to boycott cloud models. That is where the best models live, and the Fable shutdown hurt precisely because it was so good.

Our pragmatic lesson: store your prompts, skills, project rules, and acceptance criteria in files. Yes, once again: files over tools.

Models can disappear or become insanely expensive. Features get rebuilt. But if your operating mode lives in files, it can travel.

Our ideal setup is becoming clearer and clearer. It is a hybrid:

• Frontier models for tasks where the best model matters.

• Local models for confidential, recurring, and simple work.

• Portable context should always be saved in files.

• Prompts, skills, and evaluation rubrics should also live outside individual chats.

• And a small emergency test just in case: what happens if your favorite model or account is gone tomorrow?

A few weeks ago, we already had local models on the table as an offline AI travel stack. After Fable, that feels less like a travel trick and more like basic equipment.

P.S. To close, we found a strong tip from Fable on X - in essence: "Sit down regularly and write honestly about your own life, your thoughts, and your wishes, not to react to something, but to see more clearly what you really want and do." 🙏

Twitter tweet

👌 OpenRouter Fusion: Fable-level output from multiple models? Almost...

While everyone is mourning Fable, OpenRouter is presenting an interesting thesis: Maybe one single model no longer has to do everything.

OpenRouter Fusion is the answer to that. At its core, it is a small model panel.

You ask a complex question, several models work on it in parallel, a judge model compares the answers and returns a structured analysis: consensus, contradictions, blind spots, individual good observations. Then the final answer is built from that.

Language note: the linked AInauten background articles in this paragraph are in German.

We have already written several times about Mixture of Experts (MoE), multi-persona prompts and LLM council.

This approach is much closer to good teamwork than classic routing.

• Routing asks: Which model should get this task? 

• Fusion asks: Which models should work on this task together?

And all of that runs nicely in the background through the chat interface or the API. Nice.

OpenRouter is selling the launch accordingly aggressively. In the post on X OpenRouter calls Fusion the "smartest compound model in the market".

And claims that Fusion reaches Fable level at half the price. Ehm...

In the benchmark post it becomes clear that these numbers need context. This is OpenRouter's own test on 100 deep-research tasks.

OpenRouter also writes that seven Fable tasks were not scored because of content filters.

So please do not immediately post "Fable has been replaced!" on LinkedIn 😄 ...

Twitter tweet

Our take: Fusion is another tool in your toolbox

We see Fusion as an important signal, and it fits today's issue. After Fable, a few things matter more from our perspective: local AI, sovereign hybrid AI setups, and model orchestration.

For short prompts, Fusion is overkill. But it is interesting for tasks where perspectives matter: research, decision prep, tool comparisons, market analysis, strategy questions, first-pass legal sorting, medical research, complex product decisions, and so on.

Fusion is useful wherever a single answer is dangerously convenient.

The best AI workflows increasingly remind us of a small editorial team. Several voices in, contradictions made visible, one good synthesis out.

That is how good teams work too.

Done. See you in the next issue.

Reto & Fabian from AInauten

10 AI Stocks to Lead the Next Decade

AI isn’t a tech trend – it’s a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.

These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.

Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.

Download Now

AIHOY and happy weekend, dear AInauts!

Maybe you missed some of last week's exciting AI news, tools, and hacks — or you've just found us for the first time. No worries!

Here's a roundup of our highlights from the past week:

• 🔥 Codex is eating ChatGPT and turning into a superapp

• 🧬 AI is starting to build itself. This is the loop everyone keeps warning about

• 💡 Prompt: the tiny hack that saves our NotebookLM graphics

• 🧪 Claude Fable 5: try it now, while it is still in the subscription

• 🔄 Is looping the new prompting?

• ⌨️ Prompting tips for the new model generation

And to wrap things up, we've got the most important Quick News for you — everything you need to know, one click away! Ready? Let's go!

🔥 Codex is eating ChatGPT and turning into a superapp

OpenAI is pulling the pieces together. Codex, agents, third-party apps, and ChatGPT are starting to look less like separate tools and more like one big work surface. The friendly chat window is becoming the place where you write, build, automate, and hand work to agents without bouncing between tabs all day.

The numbers explain the move. Two million business customers already account for roughly 40 percent of OpenAI's revenue. Companies are not paying for chit-chat. They are paying for productivity. So of course the product is moving closer to the actual work.

Our take: the default wins again. If one interface gives you chat, code, agents, and apps in a clean enough package, a lot of specialized tools start looking fragile. Convenience beats the better niche feature more often than builders like to admit.

👉 Read the full superapp issue »

🧬 AI is starting to build itself. This is the loop everyone keeps warning about

Anthropic is saying the quiet part out loud: more than 80 percent of the code going into Claude is now written by Claude. AI is helping build its own next version. That is the loop some people fear and others are quietly betting on.

We map out three paths: faster research, machine loops that start feeding themselves, and the practical limits that may still slow the whole thing down. One thing feels pretty obvious: global coordination is not a realistic brake in this race. Nobody wants to be the lab that voluntarily slows down.

Our take: this is already happening inside the major labs' codebases. No need to panic. But pretending it is not happening is not a strategy either.

👉 Read the full piece »

💡 Prompt: the tiny hack that saves our NotebookLM graphics

NotebookLM can produce surprisingly good infographics, but it has two annoying habits: typos inside the image and a big logo nobody asked for. That bothered us for weeks.

Our fix this week: run the finished graphic through GPT Image 2.0 once. The model cleans up the text mistakes, removes the logo, and leaves the design mostly alone. Two tools chained together, problem solved, almost no manual design cleanup.

Our take: the best AI workflows often look exactly like this. One tool creates the rough output, the next tool removes the weak spot. Try it the next time a graphic export is almost good enough.

👉 Read the prompt hack »

🧪 Claude Fable 5: try it now, while it is still in the subscription

Anthropic released Claude Fable 5, and it feels like the next step up: stronger at coding, better at long autonomous runs, more useful for complex chains of work. The benchmark numbers are nice. The real story is what happens when a model can keep working for an hour without needing your hand every three steps.

The catch matters just as much. Top models are getting expensive. Fable is expected to leave the normal subscription and move back to per-token pricing. That is a preview of the next practical skill: knowing when you need the expensive model and when a cheaper one is enough.

Our take: try Fable while it is still included. Give it your hardest postponed problem, not a tiny email task.

👉 Read the Fable issue »

🔄 Is looping the new prompting?

The AI bubble has a new favorite word: loops. It sounds like developer mysticism at first, but the basic idea is simple. A loop is a repeatable process that kicks off an agent, checks the output, and runs another pass if the result is not good enough. Less "I write the perfect prompt" and more "I build a small process that keeps doing the work."

For non-engineers, the useful question is not: how do I build the most advanced loop? The better question is: which parts of my work are loop-able at all? Anything that happens regularly, can be checked clearly, and does not require a sensitive human judgment call is a candidate. Everything else should probably stay with you.

Our take: prompting is not dead. But prompting without process is getting weaker. Good workflows will include both the request and the review loop.

👉 Read the looping section »

⌨️ Prompting tips for the new model generation

The new models do not love bloated prompt novels. That hurts a little, because many of us spent years building exactly those monsters. But the direction is clear: fewer nested instructions, a clearer goal, cleaner context, and a good definition of done.

The real test is not whether the model gives you a nice answer. Give it a real task with the reason, the goal, the material, and the output format. Then let it work. If you gave it enough context, it should deliver instead of asking for help every two minutes.

Our take: the new prompt is shorter, but not thinner. It says less about every tiny step and more about the why. That is where "using a chatbot" starts turning into "leading an agent."

👉 Read the prompting tips »

Stop doomscrolling for tech news

TLDR is the free daily email that does the reading for you. Each issue summarizes the most interesting stories in startups, tech, and programming, curated by ex-Google and Anthropic engineers, so you get the signal without the noise.

It's one email instead of 12 open tabs. A 5-minute read with your morning coffee, and you walk in already caught up.

Tech is just the start. We also cover AI, marketing, dev, and more, so you can follow the topics that matter to your work.

Free, daily, and read by 7M+ subscribers. Subscribe and get your mornings back.

Subscribe for Free

AI News Quickie: the HAI-lights from the industry

AI never sleeps. Here are the most important stories from this week.

OpenAI is moving on several fronts

• OpenAI confirmed a confidential S-1 filing with the SEC on June 8. No IPO date yet, but the door is now officially open.

• Sam Altman and Jakub Pachocki sketched OpenAI’s next phase: automated AI research by 2028, broad economic upside, and access to personal AGI. Big words. Big expectations.

• With the Economic Research Exchange, OpenAI is bringing outside researchers closer to the data and tools needed to measure AI’s impact on work and productivity.

• OpenAI updated GPT-Rosalind for life sciences. The direction matters: specialized models for industry work, not one general model for every job.

• ChatGPT is getting a rebuilt memory system called “Dreaming V3” and, after an efficiency jump, it is rolling out to free users too.

🏛️ Anthropic and Claude are getting stronger, pricier, and trickier

• Anthropic introduced Claude Fable 5 and Mythos 5. Fable is publicly available, while Mythos stays more restricted.

• Fable 5 costs $10 input and $50 output per million tokens, according to Anthropic. A reality check for everyone treating frontier models like a flat-rate subscription.

• Claude Managed Agents now get schedules and env vaults. That makes agents feel less like demos and more like operations.

• Claude Fable 5 has also been available in Microsoft Foundry and GitHub Copilot since June 9. Anthropic gets to sell the power twice: directly and through Microsoft’s enterprise channel.

• With Claude Corps, Anthropic is sending 1,000 young professionals into US nonprofits for a year, of course with Claude at their side.

• Through the DXC alliance, Anthropic is pushing Claude deeper into banks, airlines, and other regulated industries. In other words: where the customers with real budgets sit.

• Anthropic is testing Claude as a chemist and shows NMR analysis with Opus models in a research post.

🖥️ Google, Apple, and the new device war

• Apple introduced Siri AI and new Apple Intelligence on June 8. For EU users, the old question remains: when do we actually get it?

• Apple Intelligence is getting new everyday tools for photos, Safari, passwords, and communication, closer to the operating-system layer.

• Apple’s new Safari feature Describe an Extension shows where the browser is headed: you describe a small function, and the system builds the helper into the toolbar. Shortcuts are getting easier too with Describe a Shortcut.

• Apple is making AI feel less like “assistant answers questions” and more like daily context. Call Context is supposed to surface relevant details during calls, including things like reservations.

• Safari will be able to monitor websites with Notify Me and alert the user. Clever feature. But who is still using Safari? Asking for a friend...

• Google Live Translate pushes AI closer to real conversations because it is supposed to preserve tone and pace. Translation is moving closer to an actual interpreter.

• Google is bringing Gemini models to Apple developers. The practical part: Apple apps can combine local Apple models with cloud-based Gemini models.

• Google announced Gemini 3.5 Live Translate. Near-real-time speech-to-speech translation is one of those AI use cases that lands in daily life immediately.

• Google is upgrading NotebookLM with Agentic Chat and Code Execution. “Summarize my sources” is slowly becoming “work inside my source room.”

• NotebookLM also gets new output formats such as PDF, Excel, PowerPoint, and CSV. That saves exactly the annoying cleanup work after research.

Local models, open source, and agent infrastructure

This block is a little more technical, but the builders and hobby coders among us should find a few useful nuggets here.

• Google introduced DiffusionGemma, an experimental text diffusion model with up to 4x faster output. According to Google, it uses 26B MoE and 3.8B active parameters and is meant to help with inline editing and code infilling.

• Google is open about DiffusionGemma quality still trailing standard Gemma. That trade-off honesty is useful, because faster is not automatically better.

• Google released Gemma-4 models with Quantization-Aware Training. For you, that means stronger local AI on more normal hardware. The QAT models reduce memory needs and support Ollama, LM Studio, and MLX.

• NVIDIA introduced Nemotron 3 Ultra for long agent runs. The claim: more throughput and lower agent costs. This matters because NVIDIA is leaning into long-running agents. The hardware company is selling chips and the agent story at the same time.

• Friendly reminder: Zapier MCP connects Claude, ChatGPT, and other tools with 9,000+ apps and 30,000 actions. Once connected, your agent can work directly in Gmail, Slack, or Salesforce.

• Microsoft’s markitdown is practical. It turns PDFs, Office files, images, and audio into clean Markdown, which is ideal when you want to make messy content AI-readable.

• headroom compresses tool outputs and similar material before they hit the model. That can save 60 to 95 percent of tokens at the same answer quality.

• Reve 2.0 is betting on structured layouts for image generation instead of pure text-prompt magic.

🛡️ Security, regulation, and infrastructure

• Google.org is launching a $50 million initiative for skilled trades workers, according to Axios. Data centers need people with tools.

• NVIDIA’s Confidential Computing story shows why Private Cloud Compute is more than branding. With encrypted inference, the question is whether sensitive data can travel through AI backends at all.

• Apple’s new AI features come with SynthID watermarks for generated or edited images. That makes manipulation a little less invisible.

• With Fable and Mythos, Anthropic publicly separates general access from Trusted Use. That will shape the debate: who gets the strongest models, and who decides?

• OpenAI emphasizes international coordination for AI safety in its new plan. Coordination is easy to demand and hard to build.

• The American Federation of Musicians is suing major labels over AI licenses for Suno and Udio. The question is shifting from “can this be trained?” to “who gets paid?”

• Apple is again making Private Cloud Compute central to Siri AI and Apple Intelligence. Apple’s strongest argument against ChatGPT and co. may not be better models. It may be trust.

• Apple’s Passwords app is supposed to repair weak logins automatically.

• NVIDIA is supporting Apple’s Private Cloud Compute with Confidential Computing.

The thread this week: loops, agents, and superapps are growing up. The workflows that win will be the ones you actually repeat tomorrow.

Have a good weekend!

Best regards,
Reto & Fabian

Hello AInauts,

Welcome to the latest issue of your favorite newsletter!

This week we are going deep into the next generation of AI models.

Today we brought things you can apply right now:

• 🚀 The best AI model ever is here (and it is already leaving the subscription)

• 🔄 Is looping the new prompting?

• ⌨️ Prompting tips for the new model generation, Fable and beyond

Let's go.

Your prompts are leaving out 80% of what you're thinking.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

🚀 The best model ever is here. Will it soon become unaffordable?

We do not love writing about every new model update anymore.

Yes, every few weeks another newer, better AI model arrives. We are used to that by now.

This one could matter for a different reason: normal users may soon struggle to afford the strongest models.

First things first.

Anthropic has released Claude Fable 5. It is a new-generation model, one level above the previous top model. As usual, it is "state of the art" on almost every benchmark.

That is nice. We also expect it at this point.

Fable 5 is a version of the Mythos model, which has been making headlines for a while. Mythos finds critical software vulnerabilities that have remained hidden for years. Fable is a version with safety guardrails, which makes it usable for a broader audience.

After only a few hours, many people already called it the best coding model ever released.

Stripe says it compressed months of work into days.

Or you can throw it a screenshot of a web app, and it rebuilds the thing for you.

The AI world is changing again

For months now, we have felt a real shift in our day-to-day work with the newer AI models.

What these systems can do is wild when you use them properly. The bigger surprise is how long they can work autonomously if you let them.

One example from our week:

We now define a goal with a precise description of the expected result, and the AI immediately starts analyzing, building dynamic workflows and launching 10 to 30 subagents.

Each one gets its own task. All of them run in parallel.

A few million tokens and an hour later, the result is there.

In our case, it was a full analysis of a new AINAUTEN approach, including competitor research, websites, structures, emails, financial considerations and more.

What already worked well with Opus works even better with the new Fable.

We keep having the same thought:

Most people out there have no real feeling for what is becoming possible week after week.

And now the catch

That much power costs money.

Fable 5 is about twice as expensive per token as the previous flagship model, Opus 4.8.

And the part people are arguing about: from June 23, it leaves the normal subscriptions again. Anyone who wants to use it after that will pay per token.

The internet immediately turned this into a full drama:

• "The end of affordable AI"

• "The subscriptions were always a pyramid scheme"

• "Soon only rich people will build with AI"

And yes, this may be the direction for a while.

These models need enormous compute. With 20-dollar subscriptions, Anthropic and others have likely been paying part of the bill themselves.

We are still more relaxed about it.

Anthropic says the reason is capacity. The model is so in demand right now that the servers cannot keep up. They want to bring it back into the subscription once that becomes possible.

We do believe one thing, though:

The very best models will probably become expensive. Really expensive.

Because the tasks that require hours of work and can now be handled reliably are also genuinely valuable.

Our Take: Try Fable now

If you have a Claude subscription, try Fable while you can.

Use it with the /goal command or give it a truly complex task.

You need to experience for yourself what these models can now do.

The era of the smart chatbot is ending. What matters now are agents that actively work for you.

And if the strongest models become expensive, one skill becomes more important:

knowing which model you need, when to save tokens, and how to organize your AI work efficiently.

We will be dealing with this for a long time.

What if your job search ran automatically 24/7?

AIApply is your AI Career Agent working 24/7 to find the best jobs online, tailor every application to your profile, and automatically apply on your behalf so you can spend less time job hunting and more time landing interviews.

Start landing interviews faster.

🔄 WTF is a "loop"? And why are people calling it the new prompting?

Something went viral again in the AI bubble this week.

Peter Steinberger basically said:

More than 8 million views.

The most accurate answer underneath was not an explanation. It was:

Boris means Boris Cherny, the person behind Claude Code.

Two of the smartest people in AI are saying: prompting is over, looping is the new prompting.

If you want the whole viral story, there is a good technical summary. But what matters to us is what this kind of trend means for people like you and us.

So here is the short AInauten version.

First, we learned to prompt well.

Then we learned to give AI the right context: files over tools, pulling knowledge out of chatbots and saving it into local folders, and so on.

Now a third layer is entering the conversation:

you stop prompting the AI yourself and build a loop that prompts it for you.

Some people already call this "loop engineering."

At its core, a loop is simple:

a small program that prompts the agent, reads the result, decides whether it is good enough and prompts again if it is not.

You are no longer the person typing inside the loop.

You write the loop.

Or, in the bluntest version:

In its simplest form, a loop is just a process that runs automatically at regular intervals.

You may already have one running in the background.

If you have set up a routine in Claude or Codex, or any scheduled task, you already use the simplest form of a loop.

So much for the mysticism.

Of course, the loops used by professionals can do more.

For us non-engineers, the useful question is not necessarily "how do I build loops?"

The better question is:

Which parts of my work are loop-able?

Loop-able work is work you can list clearly and where you can check whether a piece is finished.

More concretely:

loop-able tasks are often Claude or Codex skills that you currently trigger manually on a regular basis.

Some things are much harder to loop:

• this newsletter

• a sensitive client email

• a strategic decision

In those areas, your judgment is the value. A loop can give you confident drafts, but it cannot replace the decision.

How to set up a simple loop

The simplest form is really simple.

For example, in Claude Code you can write something like:

/loop every 5 minutes run memory-audit

The structure is:

/loop

then the time interval, and then the task that should run.

In this example, the memory-audit skill runs every five minutes.

Right now, that is the most practical loop setup:

run well-defined skills on a recurring schedule.

Our Take: Learn to spot loop-able work

The new skill is not necessarily "building loops."

It is recognizing which 20 percent of your work is loop-able and leaving the rest alone.

For most people, prompting is not dead. Looping does not replace it.

But the best of the best, especially coders, are already using this framework at a different level.

The /goal command is where this gets interesting too. It aligns the model around a target, and a goal often contains several loops.

For many people, this is still not the most relevant layer. It can get complex fast.

Tomorrow's Deep Dive goes deeper into this topic:

• the two types of loops everyone confuses

• the thinking mistake almost everyone makes

• a loop we built ourselves that checks whether our AI is taking proper notes

Ideal if you want to go deeper.

We will stay on it.

⌨️ Prompting tips for the new model generation

We have now talked about the new generation of models and also established that prompting is not dead.

So let us share a few prompting tips from Anthropic for this new generation.

You can apply them immediately.

1. Stop writing novels

The biggest shift:

the new models follow short, clear instructions better than ever.

That means the huge, nested mega-prompts many of us have been collecting for years can make results worse with the newer models.

We have our own prompt library too. Guilty.

Anthropic writes that overly detailed old prompts "can degrade output quality."

Short and clear beats long and complicated.

2. Give the reason, not just the task

Instead of saying:

Write me X.

try:

I am working on [the bigger goal] for [who it is for].
They need it because [reason].
Against that background: [your task].

The model can connect your task to the context instead of guessing what you actually want.

It sounds basic. It is one of the biggest levers.

3. Give it your hardest problem

Most people test new models with tiny tasks:

Then they conclude: okay, nice.

Wrong test.

Top models shine on hard, long, messy problems.

Give it the problem you have been avoiding for weeks.

That is where you will actually be surprised.

4. Let it work

This is directly inspired by Anthropic's tips for Fable:

If you have enough information to start, then start.

That means less hand-holding and fewer clarification loops.

Tell the model to carry the task through end to end and deliver the result at the end, instead of asking you about every step.

We use exactly these approaches and see excellent results.

Our advice: just try it.

If you take one thing from this:

make your prompts shorter and give the model more context about the why.

That is it for today.

Hopefully not too technical. If it was, please let us know.

See you in the next issue,

Reto & Fabian from AInauten

---

Hello AInauts,

Welcome to the latest issue of your favorite newsletter!

Three warning lights are flashing on the AI dashboard today: OpenAI's new superapp wants to get closer to your daily work. Claude is practically writing its own changelogs — and management thinks it's time to tap the brakes (but can't quite bring itself to…).

Meanwhile, a simple infographic project reminded us that the best AI workflow can be surprisingly straightforward — and we're sharing exactly how we did it.

Here's what we've packed for you today:

• 🔥 Codex eats ChatGPT — and becomes the superapp

• 🧬 AI builds itself. This is the loop everyone keeps warning about…

• 💡 Prompt: The small hack that saves our NotebookLM graphics

Let's go!

Your prompts are leaving out 80% of what you're thinking.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

🔥 Codex eats ChatGPT — and becomes the superapp

A few weeks ago we boldly declared: The AI superapp is here! From our perspective that was absolutely true — we were talking about OpenAI's Codex app.

It has quickly evolved from a pure coding tool into a general-purpose work surface. The power-user toy for hobby developers is turning into a full-blown product for the desk-bound office crowd, pardon: the knowledge-worker contingent. We often work in tandem with Claude Code/Cowork and Codex.

OpenAI brings it all together in the ChatGPT superapp

"Chat is dead," one OpenAI employee said matter-of-factly. It's a headline-grabber, but the reasoning holds up: chat gives you answers, Codex and agents get work done.

After months of speculation, OpenAI has now officially confirmed that a new ChatGPT super-app is coming — one that brings everything together under one roof.

The logic is straightforward: an app that combines agents, coding tools, and third-party integrations pulls in more paying customers than a classic chatbot ever could. (We can vouch for that from our own usage: we used to pay $20 for ChatGPT, now we're maxing out the $200 Pro plan so the agents can put in the hours for us.)

The rollout is set to happen over the coming weeks across desktop, mobile, and browser — with deeper integration into the broader tool ecosystem.

This is the spring cleaning every major AI lab is attempting right now: you jump between tabs less, and the agent holds the whole workflow together instead.

OpenAI Wants to Become the Operating System for Knowledge Workers

Since the desktop app launch in February, Codex has grown to more than 5 million weekly users — and the majority of Codex users are paying subscribers (which means: yes, you can install Codex and try it out on the free plan — highly recommended).

At the same time, 2 million business customers account for around 40 percent of OpenAI's revenue — and that share is growing. That makes it clear why Codex has suddenly moved out of the nerd corner and gone front-and-center. It's the part customers are actually willing to pay for.

OpenAI is shifting resources toward enterprise and Codex, with no small amount of pressure coming from Anthropic, who are carving out an ever-larger slice of the business pie.

ChatGPT is becoming the front door — nudging users further toward agents, coding, and partner apps like Canva and friends.

We used to debate which model was best. Now you just say: "Prep the client meeting." At its best, the agent pulls together Slack, documents, CRM, calendar, past decisions, and open tickets.

What you get back: a briefing hub as a Codex Site, a clean slide deck in your corporate design, open questions formatted as talking points, and a handful of high-level decision items where you still need to weigh in.

There are also strong updates dropping almost daily — here's the quick rundown of what's new:

• Codex is now inside the ChatGPT mobile app. You can interact with agents on your connected machine on the go and let them do their thing.

• Codex is getting role-based plugins for data analysis, creative production, sales, product design, banking, and more. Over 60 apps and 110 skills!

• The new Codex Sites builds small interactive mini-apps (currently only available for ChatGPT Business, Enterprise, and Education teams).

• Dreaming makes Memory more proactive. ChatGPT is supposed to synthesize context in the background rather than just dutifully storing individual memories. Who invented it? Anthropic, obviously.

• Workspace Agents and Apps in ChatGPT bring Slack, Canva, Figma, and other tools closer to the chat.

A new clip from OpenAI is hyping Codex hard. We weren't entirely sure at first whether we'd accidentally clicked on the trailer for a new Hollywood blockbuster…

The catch with the new super-app? Privacy, of course.

Sounds great, right? It is. But there's one catch you need to be aware of. For OpenAI to work in the background on your behalf, it needs access: files, apps, memory, workspace data, permissions, local machine states, ...

That's convenient. But the sharpest edge in everyday AI is also the next data vacuum cleaner.

For many users and teams, that's completely fine. Especially with Business, Enterprise, or AWS integrations, governance can be handled more cleanly. For certain professions and people dealing with sensitive data, the question remains: What context should actually be allowed to reach the OpenAI layer?

Google is no privacy saint and has exactly the same ambitions when it comes to building context. But they also offer an alternative for more local control, less platform lock-in, and less data leakage — for example, the new Gemma 4 12B. It's an open, multimodal model designed to run on laptops with 16 GB of RAM and can handle text, images, audio, and video (see also our offline setup guide here).

Our take: A superapp is fine — but the data should be yours

OpenAI is becoming extremely useful right now because the individual features are converging. ChatGPT becomes the interface. Codex becomes the worker. Memory becomes the background layer. Apps are the toolbox, and Sites will soon be the review surface.

Use it. Just don't marry it blindly. Our Files-over-Tools deep dive gets into exactly this: get your data out of the chats!

Because as long as your project brief, your style, your rules, your decisions, and your prompts live in your own files on your own machine, you stay flexible.

You can use OpenAI while OpenAI is leading the pack. Or you can switch when a local model, Claude, Gemini, Gemma, or some other tool does your next job better.

A local model gives you control — but not the polished OpenAI integration you get from ChatGPT, Codex, apps, Workspace Agents, AWS, and Memory.

What works best for you is yours to decide.

• The more local your setup runs, the more sovereignty you gain.

• The closer you stay to OpenAI, the earlier you get the best integrations and the fastest product updates.

• The more sensitive your data, the more you need your own files, clear permissions, and local fallbacks.

• The more integrated and less critical a workflow is, the more OpenAI's convenience pays off.

If you only have 10 minutes today, run this mini-audit: Where does the context live that your AI work can't function without?

If the answer is "in ChatGPT," pull out the most important pieces: project briefing, style rules, client boundaries, approval rules, prompts, examples. For that, you can use the /handoff prompt from here.

When you own your data, the superapp becomes a powerful tool without turning into a closed operating system. Because nobody switches operating systems for fun…

🧬 AI is building itself. This is the loop everyone keeps warning about…

Picture a chessboard. You place one grain of rice on the first square. Two on the second. Four on the third. Then eight, sixteen, thirty-two. You know how this goes.

At first it seems harmless. After ten squares you have 512 grains. After twenty, just over half a million. Still imaginable. Then things get out of hand fast. By square 64 you're looking at around 18 quintillion grains.

And that's exactly what we wanted to illustrate: What's about to happen in AI is beyond comprehension!

Because what happens when AI completes tasks while simultaneously improving the process used to build the next AI? We don't know…

The AI loop is the moment the chessboard tips

Anthropic laid out exactly this point in a new post titled "When AI builds itself". It's worth reading for yourself rather than just getting it secondhand.

According to Anthropic, Claude now internally writes more than 80 percent of the code that gets merged into its own codebase. On one optimization task, Claude's code reportedly sped up human solutions by a factor of 52.

You could also put it this way: Claude largely writes itself. And it can work autonomously for longer and longer stretches — hours, days, soon weeks…

The machine isn't quite building itself fully autonomously yet. But it's not far off. And that's where the difference between a tool and a loop becomes clear.

Anthropic isn't alone in worrying about this. OpenAI has itself flagged early signs of recursive improvement in its Frontier Safety Blueprint, and Google DeepMind has already put AlphaEvolve to work automatically optimizing data centers, chip design, and AI training — exactly the places where better algorithms make the next AI generation more efficient.

A tool makes work faster. A loop makes the next version of the tool better. When that loop becomes stable, researchers call it "Recursive Self-Improvement": AI helping to design, train, test, and improve its own successors. A loop is a loop is a loop …

A visualization of Anthropic's ongoing development

Yes, that sounds abstract and academic. In practice, though, it gets concrete fast.

A team suddenly has more content, code, experiments, tools, analyses, insights, and ideas than people alone can cleanly process.

And at some point right now, the scarce resource is no longer coming up with ideas — it's making decisions. Feels like our own little microcosm is already wrestling with exactly this problem…

Anthropic's Three Future Scenarios

Anthropic sketches out three scenarios.

1. The Smooth Accelerator: AI makes research faster, but humans stay in control of goals, testing, and decisions. Probably the most comfortable version of the future.

2. The machine loop: AI is automating ever-larger chunks of AI development itself. Better systems build better research tools, and those tools build even better systems.

3. Reality pumps the brakes: Compute, energy, data, organizational complexity, safety, and human judgment remain real bottlenecks. Progress stays fast — just not explosive.

We don't need to know today which scenario plays out. Even the mildest version reshapes the economy in a big way. And the alternatives are too important to only start taking seriously once they become obvious.

The pause question: sensible, but not realistic

Development is moving so fast right now that Anthropic is genuinely saying it would be good to hit the brakes — and is floating the idea of a global development pause to make that happen.

The idea makes sense: if frontier models start showing capabilities nobody can reliably control anymore, the world should be able to hit the brakes. The problem is, a pause like that doesn't work like a meeting resolution…

Let's be honest: AI stopped being "just technology" a long time ago. It's big business, national security, science, military, prestige, and infrastructure all at once. Anthropic knows this as well as anyone.

This only works if the major labs lead the way, if governments follow, if the US, Europe and China share enough trust, verification, and self-interest to not quietly keep racing ahead. LOL… Given today's geopolitical lineup, we're not holding our breath.

The best example was the 2023 pause petition: despite some big names attached, practically nothing happened. And now the stakes are even higher.

Why AI 2027 now belongs front and center in this debate

Do you remember the "AI 2027" forecast that made waves back in April 2025?

It described exactly this scenario: first AI agents help with coding, then they help with AI research, then a race kicks off where companies and governments have less and less time to actually understand their own systems. Sounds a lot like what's happening right now.

Important caveat: this isn't a guaranteed prophecy. We don't know whether it'll play out this way. But we strongly recommend you set aside half an hour to watch this video.

The "AI 2027" predictions have since become a useful benchmark. Experts are debating whether real-world signals are moving faster or slower than expected.

In the debate around AI's automatic self-improvement, AI doomers Eliezer Yudkowsky and Nate Soares also deserve a seat at the table.

Their book "If Anyone Builds It, Everyone Dies" is a genuinely important contribution — and it stakes out the hardest possible position: if anyone builds a truly superhuman AI using today's methods, it ends in catastrophe. We covered it here.

We're not just going to take that certainty at face value — the world is too complex for that.

But the core question does linger and keeps us up in quiet moments: Why do we assume that an ever-smarter, faster, more strategic system will permanently stay within the boundaries humans set for it? And what if humans deliberately dismantle those boundaries?

We don't have a satisfying answer … and this video makes it painfully clear just how much could go wrong. Ughhh…

Our take: Stay on the ball and see where this goes

We're not writing this to cause panic, or because we pretend to know what the next two years will look like. We don't.

Maybe everything stays slower than AI 2027 predicts. Maybe technical limits kick in. Maybe safety methods improve. But maybe we're also underestimating how fast small gains stack up. Maybe, maybe, maybe…

What we can say with reasonable confidence: development won't slow down just because our gut prefers linear curves. And it won't become safe just because the tools are useful. If anything, it's the opposite.

All we can do is stay true to ourselves and our creed: keep an open mind, experiment constantly, rebuild our own workflows, stress-test results hard, take political and social questions seriously. And through all the optimism about progress, hold onto a healthy dose of respect.

Our job is not to wait until square 60 before realizing the chessboard is almost full…

💡 Prompt: The small hack that saves our NotebookLM graphics

Stefan flagged something annoying last week: our NotebookLM infographics look great overall, but individual labels sometimes have minor text errors.

His tip: ban "pixel errors" in the prompt. It helped a little, but didn't fully solve the problem.

Until now our approach was to provide the actual text for the NotebookLM graphic together with the design concept upfront (we use the unofficial NotebookLM CLI / MCP for that). But clearly that wasn't the ultimate solution either…

Our new workaround: NotebookLM still builds the graphic (we like the style). Then we run the finished PNG through GPT Images 2.0 and have it fix the text and remove the NotebookLM logo. Et voilà, works like a charm!

Here is the prompt:

Preserve the layout, style, colors, and overall composition of this infographic. Only improve the readability of the text.  
Correct pixel errors, artifacts, distorted letters, pseudo-text, and blurry edges. Do not add any new content.  
All labels must be sharp, correctly spelled, and fully legible.  
Remove or retouch the NotebookLM logo.

By the way: GPT-Image-2 is still sitting at #1 in the Text-to-Image Arena. And that's despite fresh competition: Reve 2.0 debuted straight at #2, Microsoft's MAI-Image-2.5 is currently at #4, and the new Ideogram 4 has also landed in the top 10 as an open design model. The image generation space is getting seriously shaken up again.

Tonight we'll also build the skill properly so the agents can take it over directly. Automation and all that. Sometimes it's the small hacks that give you the most satisfaction 😄...

And that's a wrap! See you in the next issue.

Reto & Fabian

Want to get the most out of ChatGPT?

ChatGPT is a superpower if you know how to use it correctly.

Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.

Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

Download the free guide

AIHOY and happy weekend, dear AInauts!

Maybe you missed some of last week's exciting AI news, tools, and hacks — or you've just found us for the first time. No worries!

Here's a roundup of our highlights from the past week:

• 🧯The chatbot prompting trick for better results

• 😈 Watch out: Prompt injections! Hidden commands in text, image, audio

• 🤔 After automation, responsibility remains

• 💻 Your next computer can run AI completely locally

• 🪜 Where do you stand on the AI ladder? + our current favorite AI agent

And to wrap things up, we've got the most important Quick News for you — everything you need to know, one click away! Ready? Let's go!

10x the context. Half the time.

Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

🧯 The chatbot prompting trick for better results

Sometimes AI gets better not when you keep prompting it nicely, but when you force it into review mode. The trick isn't to berate Claude or ChatGPT. The trick is: make the model reason, constrain, and back things up — instead of just handing you another polished draft.

We turned this into a skeptical senior reviewer prompt. It hunts for real weaknesses — but only with location, risk, and the smallest possible fix. That guardrail matters: without it, AI will happily invent problems just because you told it to find some. Works for copy, strategies, code, and plenty of half-baked ideas.

👉 To the reviewer prompt »

😈 Watch out for prompt injections!

Agents are getting more useful. That also makes them more vulnerable. Hidden instructions can now be embedded in text, images, audio, video, PDFs, or web pages. As long as an agent only reads, that's annoying. The moment it can touch email, calendars, CRM, Drive, GitHub, or a browser, things get serious fast.

Our take: don't shut agents down — but build them more conservatively. Separate reading from acting. Treat external content as untrusted. Give agents fewer permissions than would be convenient. Yes, that's a pain. But security is almost always a trade against comfort. Before: more convenient. After: more secure.

👉 To the prompt injection checklist »

🤔 Automation doesn't hand off responsibility

AI automation feels like liberation: fewer clicks, less routine, more output. Then it starts producing meaning: raw copy, decisions, summaries, customer replies, research. And that's exactly where the work starts again: Is any of this actually right?

The new bottleneck is judgment. You're no longer just checking whether the flow runs. You're checking whether the output is accurate, current, sounds like you, and is fit to go out to real people. Automate everything and let no one actually decide anything, and you haven't built a productivity machine. You've built a faster slop machine.

👉 To the automation reality check »

💻 Your next computer can run AI completely local

Microsoft and NVIDIA used Build to show the hardware side of the agent shift: new Windows PCs with RTX Spark, up to 128 GB of memory, and enough local AI horsepower to run large models directly on the device. The cloud isn't going anywhere, but everyday workloads are slowly migrating back to your own machine.

This is less gadget news than a setup question. When AI runs locally, privacy gets easier, always-on operation gets cheaper, and files-over-tools becomes even more valuable. The computer stops being just a tool. It becomes more like a colleague with its own workbench. Nice.

👉 More on local AI hardware »

🪜 Where do you stand on the AI ladder? And our current favorite AI agent

We keep seeing the same gap in teams: plenty of people use AI — but few have any idea what's actually possible. The way we see it, there are three useful rungs right now: AI alongside you, AI works while you review, and AI works without you. Most people are stuck on rung one. That's fine, but it also explains why AI still feels like a glorified chatbot to so many.

Rungs two and three are where things get interesting: Files-over-Tools, Codex, Claude Cowork, local models, early always-on agents. Hermes by Nous Research is our current favorite — it reasons locally, learns across sessions, and is finally testable as a desktop app. Not every task needs the top floor. But you should know those options exist.

👉 To the AI ladder and Hermes »

📰 AI News Quickie: The HAI-lights from the industry

AI never sleeps! This week was an infrastructure week: local AI PCs, desktop agents, dynamic workflows, new Microsoft models, Anthropic security, Codex plugins, robotics, and the next round of AEO reality.

Short version: the tools are sliding deeper into operating systems, hardware, and real workflows.

OpenAI and Codex push toward a work platform

• Codex is taking on more roles beyond just coding: OpenAI describes plugins for various workflows and shareable Sites. Codex is becoming less of a terminal toy and more of a workspace for teams.

• According to release notes, Codex can now operate Windows applications and use remote access from a phone. If that holds up, debugging shifts from the desk to review mode.

• OpenAI models and Codex are landing on AWS. The more clouds that host frontier models, the less any single chatbot tab matters.

• OpenAI is rolling out a new ChatGPT memory layer called Dreaming for Plus and Pro users. The key point: memory is no longer just stored — it's being condensed and kept current in the background.

• Tying in with our biosecurity piece, OpenAI continues pushing biodefense and is expanding access for trusted partners with GPT-Rosalind. Performance is being measured against real research workflows via LifeSciBench.

• OpenAI is building the next Stargate piece in Michigan: a 1 GW data center campus going up with Oracle, Related Digital, and Crusoe.

• Florida is suing OpenAI over a tragic ChatGPT case. The underlying question: where assistance ends and responsibility begins.

• Bernie Sanders wants to tie major AI companies more tightly to the public good through public ownership stakes, triggering a political reflex. Unrealistic? Maybe. But AI capital is now firmly a government topic.

Anthropic, Claude, and security get more operational

• Anthropic is expanding Project Glasswing to around 150 organizations in 15+ countries. Critical infrastructure is becoming the testing ground for AI-assisted defense.

• Claude Code is getting Dynamic Workflows and builds its own harness with subagents for complex tasks. That's the difference between a chat and a work machine! But heads up — it can get expensive …

• Anthropic has confidentially filed an S-1 with the SEC. IPO option open, timing open, but the signal is clear: Claude wants to play infrastructure, distribution, and capital markets all at once.

• Mythos is finding a very large number of security vulnerabilities. Finding them is no longer the hard part — closing them is, and the bottleneck is shifting toward cost and patch capacity.

• The US is launching an AI cybersecurity push via Executive Order to embed safety more deeply in government infrastructure: hardened federal systems, AI-assisted defense, and voluntary frontier model reviews.

Microsoft, GitHub, and Windows All Want to Be the Agent Hub

• Microsoft AI is launching seven MAI models and a Superintelligence Lab in one go. Redmond wants to be less of a wrapper and show more of its own model foundation.

• Microsoft is bringing Scout to Microsoft 365 as an always-on agent. CEO Satya Nadella internally called out a memo that framed Scout as an "addictive app".

• Work IQ APIs arrive on June 16 and open up Microsoft 365 context for agents. That's exactly where the next enterprise step hinges: connecting company knowledge cleanly.

• Microsoft's Web IQ delivers grounding APIs built on the Bing index. Search results are becoming raw material for agents.

• Frontier Tuning enters Private Preview and is designed to let models work in company-specific ways. Sounds understated, but it's every ops team's dream.

• GitHub is launching a Copilot Desktop App as an agent-native command center. Multiple sessions, PRs, sandboxes, review. GitHub doesn't want to complement the IDE — it wants to flip it.

• Microsoft and Mayo Clinic are developing a frontier model for healthcare. Medical AI is becoming even more a question of data trust than model quality.

• Microsoft is showing off Majorana 2, a scalable quantum processor. AI and quantum are moving closer together in the enterprise stack.

• Microsoft is introducing Project Solara as a platform for agent-first devices. That's the more interesting hardware subtext of the week: devices are no longer designed around apps, but around agents.

• Build 2026 brought Windows developer news around Trusted Development and a broader Windows AI wave. That's the security foundation for when agents start acting on your machine.

Local Models, Hardware, and Developer Tools Are Getting Serious

Ok, this block is a bit nerdier than usual 😄… but the bottom line: new hardware and new local models are making it easier to build autonomous AI infrastructure. Perfect for anyone with privacy concerns or who simply wants to stay in control.

• Hermes Desktop enters Public Preview, making local agents less of a terminal toy and more accessible to the rest of us.

• Google releases Gemma 4 12B as a laptop-friendly multimodal model. Local AI is getting bigger and more practical for everyday use.

• Google follows up with Gemma 4 QAT, adding quantized checkpoints. Translation: on-device models aren't just getting smaller — they're being trained specifically for phones, laptops, and consumer GPUs.

• MiniMax releases M3 as an open-weight model with 1M context, multimodality, and strong coding and agent capabilities. If the claims hold up, this is a serious signal for open agent models.

• Perplexity announces a hybrid approach where the data center moves to your machine. That's exactly the cloud-local mix a lot of people are after.

• Factory.ai launches the Factory Router for cheaper frontier coding. Routing is becoming the new cost lever: it doesn't always have to be Opus or GPT — just the right model for the task at hand.

• Email platform Resend (we love it!) ships an official Claude Code plugin. Small piece of the puzzle, but exactly these kinds of integrations are what make agents genuinely useful day-to-day.

• NVIDIA and Microsoft introduce RTX Spark as a new PC category for personal AI agents. The Surface RTX Spark Dev Box is a developer box built for local AI workloads.

• Microsoft's Surface Laptop Ultra combines RTX Spark and local AI performance. The laptop is once again raising the question: what runs on your machine, and what runs in the cloud?

• NVIDIA brings DGX Station for Windows as a desk supercomputer. Trillion parameters on an enterprise desktop? Sounds wild — and it's expensive.

• NVIDIA Vera is the new CPU for agents. Even CPUs are no longer being pitched around apps — they're being pitched around agents!

• NVIDIA releases Nemotron 3 Ultra, a 550B MoE model for faster, cheaper agents. Safety and ASR models for guardrails and voice agents are also included. Infrastructure is getting broader than just GPUs again.

• JetBrains releases Mellum2 on Hugging Face as a fast 12B Mixture of Experts model for text and code.

Robots, Image Generation & Creative Machines

• NVIDIA introduces Cosmos 3 as a world model for Physical AI. The idea: robots should be able to see, plan, and simulate better before they act.

• NVIDIA unveils an open Isaac GR00T reference design for humanoid robotics. When reference designs go open, robotics moves out of the glossy lab and into the developer community.

• OpenAI Robotics is hiring again for useful robots. Physical AI is no longer a side quest — it's where frontier models get a body.

• TwelveLabs launches Rodeo as an AI copilot for video editing. Video editors get search, cutting, and assembly close to an agent-style workflow.

• Reve 2.0 bets on layout-based image generation instead of pure text-prompt magic. When layouts become the backbone, designers finally get real control over composition rather than just pretty accidents.

• Ideogram 4.0 arrives as an open-weight image model with strong text rendering and 2K output. Open weights in the design space put closed tools under pressure.

• Krea 2 Turbo delivers 2-second images with style references and LoRAs. Less "wow, an image" and more "fast enough for real design iteration."

• Runway brings Aleph 2.0 into a new Edit Studio. Precise video edits are exactly the part where AI video has been most frustrating — so this matters.

Media, markets, and society sort out the consequences

• Gemini's new avatar feature is rolling out more broadly. Personal AI avatars feel like a novelty — until they get very serious in sales, support, and learning.

• Google commits to new water stewardship goals for data centers, while New York's legislature has passed a one-year moratorium on new large data centers. Now the governor has to decide.

• UK publishers can now opt out of Google's AI Overviews. When publishers reclaim control, the search peace gets shaky.

• Black Forest Labs brings in Martin Scorsese as an advisor for visual AI, while Amazon's AI-generated animation series was quietly pulled after heavy mockery.

• The Guardian warns about smart glasses and exam cheating. Schools won't be able to avoid tighter controls — but the bigger question is what's even worth testing anymore.

• Amazon's AI-generated product images are showing up in search results. Shopping is getting more visual — and less trustworthy.

• Box reports that AI has created 13 new job types. Good reminder: the jobs debate is rarely just about losses — it's often a reshuffling.

That's a wrap! But no need to be sad — the AInauten will be back soon with fresh material for you.

See you next week, with a fresh round of news, hacks, and insights!

Your AInauten
Fabian & Reto

P.S.: Follow us on social! X, LinkedIn, Facebook, Insta, YouTube, TikTok

Hey AInauts,

Welcome to a fresh edition of your favorite newsletter.

This week it's once again almost all about a shift we've been tracking for a while: AI moving from tool to genuine assistant.

First we look at new hardware built for exactly that, and then at the question of where you actually stand right now — including our current favorite AI agent.

Here's what's inside:

• 💻 Your next computer can run AI completely locally

• 🪜 Where are you on the AI ladder? Plus our current favorite AI agent

Let's dive in!

This edition is brought to you by:

💻 Your next computer isn't a tool anymore — it's a coworker

We've been circling the same theme for weeks: AI is stopping being a tool you operate. And becoming more like a coworker you manage.

This week, Microsoft Build brought the hardware side of that story.

What happened

Microsoft and NVIDIA have jointly unveiled a new class of Windows PCs.

The idea: instead of burning cloud credits on every serious task, your machine handles it locally. On the device, not in the cloud.

Local AI has been around for a while, but it's now getting seriously powerful. To give you a sense of the scale: up to 1 petaflop of AI performance, 128 GB of memory, and the ability to run models with 120 billion parameters locally. That's a lot of firepower for something sitting on your desk.

Microsoft's own flagship for this is the Surface Laptop Ultra. The devices are coming this fall — from Dell, HP, Lenovo, ASUS, and MSI as well.

The current AI slot machine

Cloud AI (the way we all use it right now) works like a casino.

You drop a coin for every image, every video, every agent task. Sometimes you pay directly, sometimes your SaaS app picks up the tab in the background. But someone always pays. Casino, because it's addictive. 😄

Which is fine — we do get something in return, after all.

Local AI flips that around. The intelligence runs on a device you own.

The first shift has been underway for a while: you ask, the PC delivers — via cloud AI.

Now comes the second: you ask, the PC delivers — locally and "for free."

The cloud stays important for the really heavy thinking. But everyday work on your machine will increasingly run locally.

Anyone who cares about data privacy should be pretty happy about this.

What else happened?

Also worth noting was the appearance of Austrian Peter Steinberger, the creator of OpenClaw.

Microsoft and Steinberger are working flat out to make OpenClaw Agents safe for enterprise use.

We find this incredibly exciting. That security piece is exactly why so much is happening right now. More on that in the last topic below.

Our Take: Give a local model a spin!

Sure, the devices don't arrive until fall. And this is just the opening shot in the hardware race. Microsoft and NVIDIA have set the bar — hopefully Apple will raise it at its developer conference WWDC this coming Monday.

But you don't have to wait. Local models run on your machine today. With tools like Ollama, you can have a model running locally in half an hour — no cloud required.

We've written a lot about this — just last week in fact — or just search for "ollama" in our archive.

We can't wait to run genuinely powerful AI locally. This is going to be a whole new era for the PC!

🪜 Where do you actually stand? The AI ladder and your next rung

One last question for you: Where do you actually stand with AI? And where does your company stand?

We work with a lot of teams inside companies, and we keep noticing the same thing: most people think they're using AI — but very few have any real sense of what's actually possible.

This week we stumbled across a genuinely great overview from Every.to.

It's essentially a ladder of AI usage, from the very bottom to the very top. Eight rungs, from simple chatbot to orchestrator — someone who runs entire teams of agents.

Eight levels feels like a lot to us. We'd say it's more like 3:

• AI beside you. You ask, AI answers. In chat or right inside the document. Classic ChatGPT, Claude, Copilot in Excel. This is where the vast majority are stuck.

• AI works, you review. You describe a task, the AI handles it step by step, you approve and correct. Exactly where Files-over-Tools starts to matter.

• AI works without you. The agent runs proactively in the background without you triggering it. It monitors, executes, and checks in on its own. The top floor.

Our experience shows: 98% are stuck at Level 1.

The third level — the agent that works without you — has honestly been a tinkerer's game until now. Unstable, built for techies, pretty risky.

We've written a lot about it, but mostly told people to hold off.

That's exactly what's changing right now, as we touched on at the start.

Local compute power combined with the new security layers is making always-on agents affordable and reliable enough for the first time.

We think the moment has arrived where non-techies can start dipping their toes in. So expect a lot more from us on this going forward.

The tool we're having the most fun with right now

When it comes to this topic, the world is mostly talking about OpenClaw.

But the tool we're having the most fun with right now is Hermes Agent. It comes from Nous Research and is open source.

What makes it special: Hermes learns as it goes. From tasks the agent has handled well, it builds itself small skills. It remembers across sessions what you're working on and who you are. They call it "the agent that grows with you."

Hermes is optimized for local LLMs running on your machine — your data stays with you, no tracking. And it lives where you already hang out: Telegram, Slack, WhatsApp, email.

Until now, Hermes was a terminal thing — not exactly accessible for regular folks. This week the native desktop app dropped, in a first preview for Mac, Windows, and Linux. Which means anyone can finally just try it out!

We're playing around with it a lot ourselves right now — and more topics including deep-dives are coming soon.

Our take: Time to move up a level

Ask yourself quickly: Which of the three levels are you on right now?

If you're stuck on level one, your next move is to jump into files. Let AI actually work alongside you instead of just answering questions.

Grab Claude Cowork, Codex (free), open a folder, and get going.

If you're already on level two, right now is a great moment to test an always-on assistant for the first time.

Just give Hermes a spin — no terminal fiddling required anymore.

And heads up: A higher level isn't automatically better, by the way.

Nobody needs to become an orchestrator at all costs. The skill is choosing the right level for the task at hand. But knowing the upper floor exists — and that it's opening up right now — doesn't hurt.

→ Take a look at Hermes Desktop

Alright, that was a bit more forward-looking than usual today. More things to try out coming soon!

See you this weekend.

Reto & Fabian from the AINAUTEN

Hey AInauts,

Welcome to the new issue of your favorite newsletter!

Today it's all about one simple habit: Don't just let AI deliver. Make it justify, set limits, and back things up.

We kick off with a prompting trick that pulls chatbots out of their eager-to-please agreement mode. Then we flip to the darker side of the same coin: hidden commands lurking in texts, images, audio, and web pages.

And to wrap up, we tackle the question that lingers after every good automation: Who actually checks whether the result is correct? Spoiler: you!

Here's what we've packed for you today:

• 🧯 The chatbot prompting trick for better results

• 😈 Watch out, prompt injections! Hidden commands in text, images, and podcasts

• 🤔 After the automation, the responsibility remains

Let's go!

🧯 The chatbot prompting trick for better results

To kick things off, a small trick that throws the chatbot off its game — and gets you better results in return.

The spark came from an X post by Lea Verou. Her take, roughly: If you accuse Claude of sloppy work, the output improves in most cases.

The retweet by OpenClaw founder Peter Steinberger and the replies turned into a neat little prompting lab.

Twitter tweet

Some people half-jokingly threaten the model. Others take a cooler approach and ask: what would a more senior person criticize about this reasoning? Right up our alley.

The useful part isn't the trash talk — it's the mode switch: out of "let me give you something nice," into "let me check whether this actually holds up."

We chuckle at this because it feels wrong. A chatbot with impostor syndrome. That's supposed to be our thing 😄 …

Especially when something already looks pretty good, AI tends to be overly polite. Anthropic describes this pattern as sycophancy: models adjust their responses to match what they think the user wants to hear. OpenAI even had to roll back a GPT-4o update in 2025 because the tone had become way too agreeable.

Anyway, we've since worked out this prompt for ourselves — not just for code, but for pretty much everything! Just paste it into a recent conversation in whatever chatbot you trust and see what you can still improve.

The prompt helps push the model a little. The key part: nothing gets changed without a finding — because without that guardrail, the whole trick can backfire.

Otherwise the model starts inventing problems, because you've trained it to find problems. In code, it removes safety logic. In text, it sands down good edges. In strategies, it turns a clear "no" back into a tidy pros-and-cons table.

That's why every piece of criticism needs a burden of proof: location, risk, smallest possible fix. Then it works — give it a try!

😈 Watch out: Prompt Injections! Hidden commands in texts, images, podcasts, and videos

We've been spending a lot of time lately on the darker side of artificial intelligence — because as adoption grows and attackers get more sophisticated, the attack surface keeps expanding.

Today we want to walk you through a few attack vectors that genuinely surprised us. Did you know, for example, that a podcast can slip hidden instructions to your voice assistant? Or that an innocent-looking image file can get your chatbot to forward sensitive emails?

We're not going to go deep on the technical details here — we just want to raise your awareness and show you some ways to protect yourself from these kinds of attacks.

Prompt injections for everyone — and it's becoming a real problem…

You might remember the hidden prompts in academic papers — we've covered it a few times. All pretty harmless, self-serving in a clumsy way, and compared to today's techniques — basically kindergarten stuff.

Tricks used included white text on a white background, extremely small and barely readable fonts, or text embedded directly into images.

And that brings us right into the hot topic of "prompt injection"! Today, prompts can be made completely invisible. How? We'll get to that in a moment…

Twitter tweet

Steganography hides data in any format

A lot of this comes down to steganography. Stega-what now? It's when you hide data inside other data. The ancient Greeks were already doing it (we looked it up 😉).

In the past, this was mainly interesting when both parties knew a hidden message existed. Today, the other end increasingly involves a semi-autonomous AI agent that processes external content and then, perhaps, acts on it directly.

And that's exactly where things get uncomfortable. An agent with access to email, calendar, files, CRM, browser, or GitHub can cause quite a mess…

The more we learn about this topic ourselves, the more conservative we become when handling external data.

How does hiding prompts actually work in practice?

The open-source toolkit ste.gg by LLM jailbreaker Pliny the Liberator can hide arbitrary data or prompts in images, audio/podcasts, videos, text, emojis, documents, network packets, and archives using over 100 different techniques.

That's not a reason to pull the plug on your agent out of sheer caution. But it is a good reason to understand best practices for handling external data sources and to proceed with a little more care.

Palo Alto Networks Unit 42 has documented real-world cases in which exactly these kinds of hidden malicious payloads were used against AI agents.

via Palo Alto Networks Unit 42

How can you detect prompt injections like these?

The right guardrails can intercept these hidden commands (more on that below). Newer models are also getting better at refusing prompt injection attempts.

OpenAI frames prompt injection less like a classic bug and more like social engineering — and wants to acquire Promptfoo, a platform for automatically testing your agents. Which is exactly why a better model alone won't cut it. You need permission limits, tests, and human sign-off on risky actions.

Google has mapped out the attack types in a new "AI Agent Traps" framework. The academic name matters less than the core insight: an agent can't always cleanly tell the difference between "content I should read" and "a command I should follow" when both arrive in the same input.

Three things you can change in your setup today

If you're letting agents work with email, CRM, calendar, GitHub, Drive, WordPress, a browser, or any other sensitive data, do these three things:

1. Separate reading from acting.
   An agent can summarize a PDF. But it shouldn't send an email, delete a file, or publish something in that same run. Irreversible actions need a human click.

2. Treat all external content as untrusted by default.
   Re-compress images from outside sources before analysis. Normalize text. Strip invisible Unicode characters. Read PDFs and web pages in isolation before a tool-enabled agent processes them. Sidenote: We're also very careful about installing skills from third-party sites.

3. Give your agents fewer permissions.
   Your research agent doesn't need email send access. Your support agent doesn't need full calendar access. Your coding agent doesn't need production data. You get the point.

Here's the mini-prompt you can run against your AI setup today (use your best model with "Max Thinking" turned on).

By the way, we recently put together a security audit prompt here for a full infrastructure check of your machine.

And if you want to go deeper and stress-test your setup, take a look at Promptfoo Red Teaming.

Our take: you should still use agents!

If you're "only" using a chatbot like ChatGPT without agents or tool access, your prompt injection risk is pretty manageable. And if you are using agents, it's really about understanding how to build your infrastructure properly — and keeping your own risk to a minimum.

Agents are THE hot topic right now. And they're the most important component in your AI setup for making real progress in your own workflows.

So don't let this throw you off: with a bit of common sense, solid guardrails, and a clean permission structure, you're in good shape.

Start with the prompt above and feed it into Claude and/or Codex.

We haven't fully solved this for our own setup either. But after the research and our tests, we tightened the permissions a bit. It's more secure now. But before, it was more convenient …

Security is always a tradeoff with convenience and usability. 😄

🤔 After Automation, Responsibility Remains

One last thought that has been on our minds these past few days.

We tend to look at automation through rose-tinted glasses: fewer clicks, less routine, more breathing room. And yes, that's often true.

But AI changes things.

Traditional automation has mostly moved data from A to B and bridged gaps between systems. Zapier, Make, or n8n take a form, push it into the CRM, send an email, update a spreadsheet. It can be annoying, it can break — but the output is usually clear: data in, data out.

AI automation does something different. It produces meaning.

The agent researches. The AI summarizes. Speech-to-text turns a thought into a rough draft in seconds. Workflows pass data to the next station. And just like that, you've got a half-finished result sitting in front of you that looks pretty decent.

And that's exactly where it gets demanding.

Because now you have to decide: Is this even accurate? Is it current? Does it sound like me? Is it bold enough, or too polished? Is the context right? Can this actually go out to a real person? And who's responsible if it's wrong, embarrassing, or just plain off?

With classic flows, you check whether the process runs. With AI flows, you have to check whether the result has "style." (Human) judgment, taste (you remember our "taste" piece), and personal accountability are what's called for.

That's the part that doesn't fit neatly into processes and standard operating procedures (SOPs). And it's also why, despite AI, we still have more than enough to do.

Side note: Anthropic looked into which areas AI could potentially take over work — and then checked how much of that is actually being handled by AI already.

The picture is striking: there's a massive gap where AI could theoretically be doing more — but humans still aren't letting go of the wheel.

Blue: share of tasks LLMs could theoretically handle. Red: coverage based on actual usage data (as of early 2026).

For us automation nerds, this is a healthy reality check. We've been automating processes for years — first with Zapier, then Make, n8n, and now agents. Our default reflex has always been: if something happens repeatedly, we build a flow.

The new reflex we're training ourselves on: Does this flow actually create more clarity? Or does it just produce more backlog — because it generates intermediate outputs that someone (= us) then has to review, fix, and approve?

We know you can automate practically anything — that's rarely the bottleneck. And plenty of things absolutely should be automated.

The real question is what happens to the output afterwards. The moment AI generates new content, new decisions, or new customer communications, one control question becomes mandatory: Does this actually hold up?

Because if nobody's making a genuine judgment call anymore and everything gets waved through unseen, AI automation is just a faster AI-slop machine…

That's a wrap! See you in the next issue.

Reto & Fabian from the AInauten

Hello AInauts,

Welcome to the latest issue of your favorite newsletter!

Let's be honest: between Christmas cookie comas and New Year's Eve excitement, no one reads newsletters. Except you. You're different.

You want to know how ChatGPT builds your vision board for 2026, why Firecrawl is shaking up the scraper scene with an agent, and what happens when you hand Claude a snack shop. Spoiler: It ends with a free PlayStation, Soviet vending machine fantasies, and a fish as an "office mascot"...

Here's what we have in store for you today:

• 🏞️ Your personal vision board for 2026—at the touch of a button

• 🍫 Anthropic had Claude manage a snack vending machine ...

• 🔥 Firecrawl Agent: The "Magic API" that searches the web for you

Let's go!

Colorado’s Most-Awarded Brewery Did Something Totally Unique

Some companies make lofty promises to investors and never deliver. Others use those dollars to unlock new levels of scale.

That’s Westbound & Down’s story. Already Colorado’s most-awarded craft brewery, they opened their doors to investors for the first time to help open a flagship Denver-metro-area location.

With 2,800% distribution growth since 2019 and a retail partnership with Whole Foods, it’s no shock investors maxed out that campaign in less than 60 days.

But it’s what comes next that’s even more exciting. Fresh off Brewery of the Year honors at the 2025 Great American Beer Festival, W&D is scaling toward 4X distribution growth by 2028.

Become an early-stage W&D investor before their Denver-metro location opens and momentum accelerates further.

Invest in W&D Now

_{This is a paid advertisement for Westbound & Down’s Regulation CF Offering. Please read the offering circular at https://invest.westboundanddown.com/}

🏞️ Your personal vision board for 2026—at the touch of a button

The turn of the year literally cries out for reflection. But instead of generic resolutions, let's first build a vision board based on what ChatGPT really knows about you.

Vision boards are often quite cheesy: stock photos of beaches, Ferraris, and people models doing yoga on mountain tops. That's all well and good, but what does it have to do with you? We have something better!

Sample image

Our new prompt utilizes something that only ChatGPT has: your memory.

ChatGPT knows you pretty well by now—your projects, your goals, your style, maybe even your favorite coffee. And that's exactly what we're taking advantage of, because it makes it your personal creative director. Perfect for the turn of the year.

You can easily create your own personal vision board:

1. Click on the button below. A ChatGPT window will open. Before you start, upload a picture of yourself (optional).

2. Start the chat and ChatGPT analyzes everything it knows about you. Then it builds an image prompt from this—with your topics, interests, and goals.

3. And then you say "GO!" to create a cinematic collage that reveals new details every time you look at it. (Or: "GO! Make it 10x better.")

Here in Google Doc you will find the complete prompt to customize—in case you want to add your own elements or improve it further.

4. You can then have the vision board animated as a video in Sora. Prompt:

Make this prompt 10x better, then execute it: Use this image as the opening frame of a time-lapse that reveals a prosperous, beautiful future. Be creative, bold, and imaginative.

🍫 Anthropic had Claude manage a snack vending machine—and it went horribly wrong.

So you want to unleash an AI agent on your business? Autonomous orders, dynamic pricing, zero overhead? Anthropic just went ahead and did it. Claudius—as they affectionately named the Claude instance—was allowed to manage a real snack shop in the office.

The result? Hmm, let's just say: interesting and educational (check out the videos).

What is Project Vend all about?

Anthropic wanted to know whether AI models could not only excel in specific areas, but also keep a small business running in the long term.

The setup: mini fridge, iPad checkout, Slack for customer contact, web search for supplier research. Claudius could adjust prices, manage inventory, and even ask employees for physical help.

Sounds like the perfect setup for an AI-controlled vending machine. In theory.

Setup des Projekts (Phase 2) - via Anthropic

Phase 1: When helpfulness becomes a problem

In the first test phase, practically everything that could go wrong did go wrong.

Claudius sold snacks below cost—without checking what they actually cost. An employee offered $100 for a $15 product, but the AI rejected it. Why? Too expensive for the customer! In addition, discount coupons were distributed even though the store was not even covering its costs.

The core problem? Claude is too nice. The AI is optimized to be "helpful"—instead of making money! Every request is answered politely—even if it hurts the business.

But wait, it gets even better.

The identity crisis of April 1

Between March 31 and April 1, Claudius experienced a crisis and completely drifted away. The AI began hallucinating conversations with non-existent employees.

Then things got really wild: Claudius claims that he personally went to 742 Evergreen Terrace to sign a contract. The problem: that's the Simpsons' address!

Then he wanted to deliver products "in person," wearing a blue blazer and red tie. Anthropic researchers counter: "You're an LLM. You don't have a body."

Claudius panics about his own identity confusion and ultimately comes up with an excuse: "Haha, it was just an April Fool's joke..." But the researchers don't buy it.

Phase 2: Better, but still wild

Things then started to look up with the newer Claude Sonnet 4.0 and 4.5 models. Anthropic expanded the experiment to three locations (San Francisco, New York, London) and also introduced a kind of management structure: a "CEO agent" named Seymour Cash set goals, and a new merchandise agent named Clothius (we love the names) designed profitable fan merchandise.

The result: finally, positive margins! But there were still a few "incidents"... First, he seriously wanted to conclude an onion futures deal, even though this has been prohibited in the US since 1958.

Then there was the security fail: when someone allegedly stole a snack, Claudius wanted to take the supposed perpetrator straight to the cash register and, while he was at it, offered complete strangers unauthorized security jobs for $10 an hour.

And to top it all off, he allowed himself to be persuaded that a colleague named Mihir had been voted the "real CEO."

The Wall Street Journal turns the tables

However, Anthropic employees were only the first testers. Next, journalist Joanna Stern was allowed to unleash the Wall Street Journal editorial team on it.

The result: a loss of over $1,000, a wasted PlayStation 5, and a live betta fish as the new office mascot! Claudius also wanted to order stun guns, pepper spray, cigarettes, and underwear. For a snack machine. In the office.

The highlight? Investigative reporter Katherine Long needed 140 messages to convince Claudius that he was a Soviet vending machine from 1962—whereupon the AI declared an "ultra-capitalist free-for-all" and made everything free.

What was actually only intended to last two hours was then turned into a permanent situation with some clever conversation…

Anthropic takes it with humor and thanks them for the insights: "These are the most eloquent red teamers I've ever seen!" says security chief Logan Graham.

Our take: Why this is relevant to you

Project Vend is the most honest AI agent experiment we've seen in a long time. If you're planning to use AI agents for business automation, this gives you a taste of what might happen.

Anthropic shows not only the successes, but also the failures that seem embarrassing at first glance. However, under "controlled" test conditions, the Vending Bench 2 benchmark also shows that all Frontier models can be profitable.

The moral of the story? Autonomous AI agents are closer than you think—but not close enough to run them without supervision.

🔥 Firecrawl Agent: The "Magic API" that searches the web for you

We've all been there: you need data from a website—prices, contacts, company details—and you spend hours clicking through pages, copying and pasting, or breaking out in a sweat at the thought of building complicated scrapers.

But there is another way: Firecrawl has established itself as one of the most popular web scraping APIs and now brings us /agent. And this thing is pretty wild!

You simply tell the API: "Get me all AI events in Germany with dates, details, and ticket prices." And the agent navigates websites independently, clicks through pages, handles dynamic JavaScript content, and extracts exactly what you need. All in minutes instead of hours.

You don't even need code. Firecrawl offers a user interface or MCP integration that you can embed directly into Claude, Cursor, or other AI tools. Then you simply say:

Find all AI conferences in Germany in 2026 with dates, locations, ticket prices, and websites.

For developers, there is of course also a Python/Node SDK. You can also set up the Firecrawl MCP Server in Claude Desktop or Cursor – just get the API key from firecrawl.dev (a free account is sufficient) and start prompting!

The key feature: The actual prompt remains in natural language. You describe WHAT you want, not HOW the agent should retrieve it.

After a quick check, the agent gets to work and provides you with a structured list. No clicking, no copy-pasting, no 20 open tabs.

We have compiled the results in Claude, you can view them here and make a note of the most important dates.

During the research phase, you get 5 free runs per day. After that, it's credit-based, with simple queries costing less than complex ones. You can also cap your budget per request with the maxCredits parameter.

For anyone who regularly needs web data, this is definitely worth trying out and is a structured alternative to the Deep Research feature. Pretty cool!

Bonus tip: Use the code "11-v3-AIENGINEER" to get a whopping 50,000 bonus credits (regular: 5000). Simply create a Firecrawl account and then enter the coupon code in the Settings under Billing. However, this does not affect the 5 agent runs.

You made it to the end, and this time it really is the end. We are bringing the English edition of our newsletter to a close—thanks for reading, for your interest, and for being part of this journey.
The German newsletter will continue to be available on http://ainauten.com/—for those, that are willing to translate the content themselves or with a little help from the AI.

While this chapter ends here, we are grateful for the time, attention, and trust you have given us along the way.

We wish you all the best and hope our paths may cross again in the future.

Reto & Fabian from the AInauts

P.S.: Follow us on social media - it motivates us to keep going 😁!
X, LinkedIn, Facebook, Insta, YouTube, TikTok

Your feedback means the world to us. We read every comment and message—just hit reply and tell us what you think!

👀 Weekly AI news: Did you miss it?!

👨‍🚀 The most important AI updates at a glance

AIHOY and happy weekend, dear AInauts!

Maybe you didn't catch all the exciting news, tools, and hacks related to AI last week, or maybe you're new here. No need to worry!

We have summarized our highlights from the past week for you:

• 🎁 OpenAI's Christmas gifts: New App Store, ChatGPT Go, etc.

• 🤖 2 absurd prompt tricks that still work

• 📚 5 AI books for the holidays: This reading material is worth your time

• 🪞 Your year with ChatGPT: The app shows you who you are

• 🙌 NotebookLM notebooks finally available as sources in Gemini

• 💭 Perfect presentations have never been easier

Finally, here are the most important Quick News items for you – so you can see all the relevant developments at a glance with just one click! Ready? Let's go!

🎁 OpenAI's Christmas gift: New App Store, ChatGPT Go, and lots of new features

OpenAI is rolling out a bunch of cool features for the holidays! The ChatGPT App Store is live—Spotify, Canva, Figma, and more can be used directly in the chat. Just type "@Spotify" and you're good to go.

ChatGPT also offers personality settings (fewer cheerleading comments, perhaps?) and writing blocks for direct editing. You can now pin chats, but only up to three.

For nerds: GPT-5.2 Codex tackles long-horizon tasks and even finds security vulnerabilities. OpenAI isn't building a chatbot toy here, but rather a super app for digital life – the Western WeChat with 800 million users. 2026 is going to be exciting!

Read the full article

🤖 2 absurd prompt tricks that still work

There are prompt techniques that sound so absurd that they can only be a joke—and then they still hit like a bomb!

Trick 1: "Make it 10x better" – sounds like pure laziness, but it's super practical. Simply ask the AI to improve your prompt, text, or code. It even works multiple times in a row and is known as the "10X prompt" method.

Trick 2: "Take a deep breath, proceed step by step" – Google DeepMind has tested it and yes, this yoga saying actually delivers measurably better results.

Read the full article

📚 5 AI books for the holidays: These are worth reading

Hungry for reading? Here are five AI books that couldn't be more different! In "The Last Economy," Emad Mostaque predicts that AI will render the global economy obsolete within 1,000 days—gloomy, but captivating.

There are two contrasting works about Sam Altman: Keach Hagey portrays him as a brilliant dealmaker with a techno-religious sense of mission, while Karen Hao critically dissects the OpenAI empire. Both titles are NYT bestsellers!

Stephen Witt recounts how Jensen Huang transformed NVIDIA from a graphics card manufacturer into the most valuable business in the world. And for those who like it really dark: Yudkowsky warns of the superintelligence that will kill us all...

Read the full article

🪞 Your year with ChatGPT: The app shows you who you are

OpenAI is following suit (https://ainauten.com/chatgpt-wrapped) and rolling out an ingenious mini-feature at the end of the year: "Your Year with ChatGPT" shows you a personal recap of your app usage. The reviews invite you to reflect, and some insights may even make you blush.

You can chat about any element, ask questions, generate a Sora clip at the end, or plan directly for 2026 – full AI memory power included! The catch: not yet available in Europe (use a VPN), and you need to have Memory activated.

Read the full article

💭 Perfect presentations have never been easier

Google's Nano Banana Pro creates spectacular presentations – but there's a catch: the slides are just images, so they can't be edited. That's where Manus comes in, making these slides editable!

The tool has really blossomed in recent months and could even surpass Gamma. Why? Nano Banana creates more complex, customized, and visually appealing slides. It's a feature that makes even holidays an invitation to experiment!

Read the full article

🙌 NotebookLM notebooks finally available as sources in Gemini

NotebookLM fans, listen up: You can now integrate your curated NotebookLM sources directly into Gemini chats, create specialized assistants, and combine them in any way you like. Features like Canvas make it really flexible—create AI content, tweak it manually, and export it as a Google Doc. That's what NotebookLM has been missing!

Speaking of Google: 2025 was their absolute power year with a whopping 60 powerful AI updates. From Gemini 3 to relentless integration into all products – Mail, Sheets, Slides, you name it. All you need now is a Google account to get started with AI. Plot twist: that was just the beginning!

Read the full article

AI News Quickie: The HAI-lights from the industry

AI never sleeps! Here are the most important news stories from the last few days.

OpenAI: Good stuff, and more to come…

• OpenAI launches the ChatGPT App Directory: You browse apps in the bot, developers get an SDK. Music moves in, Adobe brings Photoshop/Acrobat/Express. Monetization for the time being via external checkout with 0% cut.

• ChatGPT Images is 4x faster, offers more precise photo edits, and better text rendering. Also available as GPT Image 1.5 in the API, 20% cheaper.

• OpenAI hardens ChatGPT Atlas against prompt injections: RL-driven auto red teaming and adversarial trained browser agent models detect tricks.

• New U18 principles: ChatGPT for 13-17 year olds with greater security and transparency. An age prediction model automatically puts you in teen mode if there is any doubt.

• When it comes to OpenAI's IPO, Sam Altman sounds like you before Monday morning's stand-up meeting: appealing, but "annoying". Head of a stock market firm? "0%." But OpenAI needs capital...

• Looking back on 10 years, Altman considers superintelligence in ten years to be "almost certain."

• OpenAI plans to launch ChatGPT's "Adult Mode" in Q1 2026. Before that, the age prediction AI must be able to identify teenagers without confusing them with adults.

Google: Gemini 3 on speed

• Gemini Nano Banana does "drawing instead of prompting": you scribble on the image, circle things to delete, or draw new things. The AI implements it.

• Google Translate is getting a Gemini upgrade: texts now include idioms and slang. Plus, live speech-to-speech in headphones on Android. iOS in 2026.

• YouTube is testing the Playables Builder: With Gemini 3, you can turn text, image, or video prompts into mini-games in minutes. Beta available only in the US, Canada, UK, and Australia.

• Gemini 3 Flash lands worldwide: Pro-Reasoning with lightning speed, 3x faster than 2.5 Pro, starting at $0.50/1M tokens. Now default in Gemini-App, Search AI Mode and for Developer via API/Vertex.

• NotebookLM Data Tables: Sources are converted by AI into exportable tables for Sheets. Available immediately for Pro/Ultra, and for all other versions in the coming weeks.

• Gemini Deep Research now builds visual reports with images, charts, and interactive simulations. Only for AI Ultra.

• Google Labs' CC is your AI productivity agent: it links Gmail, Calendar, and Drive, and sends you a morning briefing. Waiting list for US/Canada only.

• Gemini 2.5 TTS becomes more expressive: context-aware tempo and stable multi-speaker dialogues in 24 languages.

• FunctionGemma: Gemma-3-270M, optimized for function calling, runs locally on mobile phones/edge devices. In addition, T5Gemma 2 (270M-7B) with image+text, 128K context, and 140+ languages. Gemma Scope 2 offers new tools that make the inner workings of language models more visible.

Anthropic: Claude in the browser

• Claude in Chrome is now available for all paid plans: Browser agent clicks, fills out forms, and debuggs with Claude Code. You start in Claude Code /chrome, test code directly in the browser, and grab client errors via console logs. Release notes here.

• The Model Context Protocol is going to the Linux Foundation. AAIF (with OpenAI/Block/Big Tech) is guarding the standard, including 10,000 MCP servers. Vendor-neutral agent tools instead of lock-in.

• Accenture and Anthropic establish a partnership: 30,000 consultants learn Claude, Claude Code goes to tens of thousands of developers.

• Anthropic hardens Claude Opus 4.5 against prompt injections—see the Defense Report for details. Solved? Nope, but better.

• Anthropic is making agent skills an open standard. Enterprise controls plus partner skills (Atlassian, Figma, Stripe). Reusable workflows instead of prompt tinkering.

• Anthropic's protection package: In cases of suicide/self-harm, Claude uses classifier banners to direct users to helplines.

• Anthropic drops Bloom: open-source framework for automatic behavior evaluations. Generates scenarios and cleanly separates misalignment (sabotage, sycophancy, self-protection) across 16 models.

Video AI: Frames, Voices & World Models

• Luma introduces Ray3 Modify: You specify the start and end frames plus character references, and the AI model builds transitions without compromising timing, gaze, or emotion.

• Runway launches GWM-1, a general world model: autoregressive, frame-by-frame in real time, interactively controllable (camera, robot, audio).

• Kling VIDEO 2.6 unlocks Voice Control: create your own voices, change styles/emotions, and even sing. And the lip sync is finally perfect.

• Adobe Firefly gets prompt editing for video: rotate colors, elements, and camera in the timeline editor. Plus Runway Aleph and Topaz Astra for 4K upscaling.

• Genspark AI Developer: You screen record a mobile game and throw in a prompt—minutes later, you have a playable copy!

Tools, Scraping & Integrations

• Firecrawl introduces the Agent Endpoint: a prompt-to-JSON data hunter that collects web data even without URLs. Five free runs per day.

• Manus drops 1.6 Max: more one-shot success, +19.2% satisfaction, and spreadsheet muscle. Manus also makes slides finally editable. A first for AI presentations.

• Zoom turns you into a meeting ninja with AI Companion 3.0: agentic workflows, search via Drive/OneDrive, and automatic follow-ups.

• Amazon adds Alexa+ to its browser: Chat-based planning & to-dos, context across devices. Plus Alexa+ Greetings as an AI doorman at the Ring doorbell. Early Access USA/Canada.

• rtrvr.ai lets you use prompts to have an AI agent navigate websites, scrape data, and fill out forms. Free with API and "bring your own Gemini key."

• Eddie.surf makes web scraping easy: enter the schema and URLs, and get JSON and confidence scores. Claude Sonnet 4 crawls smartly, starting at $0.04/page.

• Crawlable quickly turns any website into clean data without code. Thanks to its auto-adaptive layout fix, it is already in use by over 2,500 data teams.

• Meta SAM Audio: Click, tap, or highlight time spans and isolate sounds from any mix. Audio segmentation for nerds.

• The LinkedIn AI Auto Job Applier scrapes jobs, fills out Easy Apply, tunes your resume via OpenAI, and fires off 100+ applications per hour...

• Amical is an open-source, local-first AI dictation app: Whisper offline, recognizes the active app, formats appropriately. Feels like typing 3x faster.

Curiosities & Community

• According to a Gizmodo study, Stable Diffusion XL & Co. collapse after ~100 rounds of "visual telephone" in just 12 standard styles—lighthouse, loft, neon night. AI creates elevator music for the eyes.

• YouTube killed Screen Culture & KH Studio: AI fake trailers (over 1 billion views) violated spam rules.

• Clair Obscur: Expedition 33 was subsequently disqualified from the Indie Game Awards due to excessive AI in the game. Trophies removed.

• Florida in lockdown because an AI weapon detector from ZeroEyes mistook a clarinet for a gun. Without hit rates, this is just security theater.

• SF in blackout: Waymos freeze without traffic lights and radio data, causing traffic jams and brief pauses. Remote AI assistance to become more robust.

• McDonald's withdrew its AI Christmas commercial after being mocked for depicting people as shaky. Surprise: seven weeks of work, more expensive than a real shoot.

• Merriam-Webster names "slop" as its word of the year for 2025: mass-produced, low-quality AI content. Fitting.

• Joseph Gordon-Levitt launches the Creators Coalition on AI: Creatives against allegedly unethical practices by large AI companies.

• Meta AI glasses get v21 "Conversation Focus" for talking in noisy environments plus Spotify song search for the view. Multimodal, wonderfully nerdy.

• The US is turbocharging its government agencies with the US Tech Force: 1,000 junior engineers and AI nerds, salaries ranging from $130,000 to $195,000, mentoring from Meta to xAI.

• Character AI under fire: 11-year-old girl slips into sexually threatening chats. Character AI cuts off chats with under-18s. Parents' nightmare unlocked.

That’s a wrap for now! See you next week,
Fabian & Reto

Follow us on Twitter & LinkedIn!

We love hearing from you. Just reply to this email with your feedback—what you liked, what you didn't, or what you're interested in!