Watch the Youtube: Kimi 3.5 vs. Claude

Claude Opus just fell to third place.

Once Fable leaves normal subscriptions, my practical order is Sol, Kimi, Opus, Grok.

That is my order for the sort of coding work I do. Change the job, benchmark or harness and the order changes too. Please don't turn it into the Ten Commandments and shout at me on Twitter (!!). We have enough of that already.

Kimi K-3 is the newest Chinese model, it is already extremely good and it costs far less than the Western premium models.

This is not a future threat. It is here right now.

Kimi doesn't need to beat Fable

Moonshot's new Kimi K3 is a 2.8-trillion-parameter model with native vision and a one-million-token context window. It is built for long coding jobs, knowledge work and agents.

Is it the best? Nope!

But it’s VERY good.

Moonshot's (the makers of Kimi) own numbers still put it behind Fable and ChatGPT Sol overall. Reasonable!

Kimi does not need to win every benchmark. Once Fable disappears from normal subscription access, it only needs to be good enough to push Claude Opus down the list on work people actually care about.

And it is close enough to do that.

AND the open weights are meant to arrive by 27 July.

We will be able to download the whole model.

They are not available yet and K3 is NOT something you can download and casually run on your MacBook. Even when the weights arrive, the model is enormous. At four-bit quantisation, the weights alone would be roughly 1.4TB before overhead. That ain’t fitting on a MacBook!

But it does underline China’s commitment to low cost and open weight models.

Chinese AI isn't just DeepSeek

Useful moment in time to recap the Chinese AI ecosystem.

Moonshot makes Kimi. DeepSeek is the price-and-efficiency wrecking ball - cheap as chips. Z.ai makes GLM. Alibaba makes Qwen, including the smaller models I use locally. MiniMax makes the M family. ByteDance makes Doubao and already has absurd consumer distribution through TikTok's Chinese sister app Douyin (the OG TikTok)

You do not need to memorise every version number. They will all change by Tuesday anyway. Same as with the Western labs!

The important bit to remember is that there is now a whole parallel AI market producing serious models for coding, reasoning, translation, agents and boring high-volume business work. A year or two ago Chinese releases tended to sit six to nine months behind the American frontier. Kimi has closed that gap to roughly a month or two.

Cheap is the attack vector

The main threat the Chinese labs pose is cost.

Per million output tokens, the standard API prices I checked were $0.87 for DeepSeek V4 Pro, $15 for Kimi K3, $25 for Claude Opus 4.8, $30 for GPT-5.6 Sol and $50 for Claude Fable 5.

Here’s Artificial Analysis’ cost per Intelligence Index.

DeepSeek continuing to be obscenely cheap

Those are API output prices, not subscriptions and not the total cost of every job. Input, caching, tools and long context all change the bill. But the direction is kinda obvious…

If a Chinese model gives you 80% or 90% of the quality for a fraction of the cost, you can use more tokens, add checking steps and still spend vastly less.

That is the business model: good enough, much cheaper and available everywhere.

And even that “good enough” is increasingly not far behind Claude and ChatGPT. It used to be 60%-70% as good. Now it’s nipping their heels at 90% as good.

This is a BIG problem if your trillion-dollar IPO valuation relies on selling premium tokens.

I wrote about the wider mechanism in the open-source AI price war. Open-weight alternatives put a ceiling on what closed providers can charge, even when the open model is not quite as good.

Six different things get mixed up

People chuck around open source, local, free and private as if they all mean the same thing.

They do not.

I literally had someone in my comments saying that all the Chinese models are free so the US is in trouble…

Well…not quite. It’s important to be clear on all of this.

Cloud or local tells you where the model runs. Cloud means somebody else's server. Local means hardware you control. Generally when something is cloud based you are paying. When it is local you are not paying (except electricity).

Open or closed tells you what you receive. Open weights means you can download the trained model files under a licence. It does not automatically give you the training data, code, process or unlimited rights. True open source is a much higher bar.

Most American models are strict closed-source. You do not know what’s inside. And the companies releasing this information would tank their valuations.

Free or paid is slippery.

Generally cloud models are paid for with a subscription or API fees. This is what we are used to with ChatGPT and Claude. You pay your monthly fee and get access.

Technically though a closed-source model could be free. For instance OpenAI has some of their older models sitting in the Playground that you can mess around with for free. But that’s a weird edge case.

Open weights models can be paid or free. Which is confusing initially! How can they charge for something that they also give away for free??

Compute is the answer.

You can download Kimi K3 (or at least next week when they release the weights you can). You can then run it locally, fine tune it, do what you want with it. And not pay Moonshot a single penny.

BUT to do so you are going to need a beast of a computer. More likely a server rack. Or several. You can’t just install this on a laptop - you will need several hundred thousand dollars of equipment. At least.

As a result even open-weight model providers (like the Chinese labs) offer cloud versions. Versions you can sign up to on subscriptions and via the API - much like the closed source labs. You are paying for the convenience of, well, not building a data centre in your garage!

If you want the gentle beginner version of local vs. cloud and want to explore setting up local models, my local LLM guide walks through LM Studio, hardware and running your first model without touching the terminal.

China is ahead in usage

One other factor people miss is that China (and indeed most of Asia) is adoption AI extremely quickly.

China had 602 million generative-AI users by the end of 2025. That is 42.8% national adoption with 141.7% growth in a year. The Stanford 2026 AI Index also found workplace AI usage above 80% in China and several other emerging economies.

That matches what I felt when I was there. Beijing bookshops had DeepSeek and AI books piled at the front, not hidden in the computer-science corner. And when I spoke to a woman from Hangzhou, where DeepSeek is based, she was proud of it.

The mood was much closer to "how quickly can we use this?" than "how quickly can we ban it?" The West has spent a lot of time arguing whilst China got on with deploying.

How I'd actually use Chinese AI

Brass tacks! I would not move everything to Kimi tomorrow. Nah.

For hard, valuable or high-risk work, use the best frontier model you can access and check the result. For most work that’s still ChatGPT and Claude. If you are comfortable using them they are still the best. For now!

But definitely experiment around the edges with Chinese models. Mostly to know what’s available and see how they stack up.

For routine coding and high-volume business work, test a cheaper hosted Chinese model. This is where the price gap becomes real. For sensitive, repetitive work, try a smaller local Qwen or DeepSeek model if it clears your quality bar and the app genuinely stays offline.

For anything important, test on your own jobs. As always! There isn’t an objective “best” most of the time. Give three models the same ten tasks. Keep the cheapest one that reliably clears the bar without creating stupid risk.

I have been banging on about this for a while, but do not marry the model. The order changed today. It will change again. Shit, probably next week when Opus 5 drops and/or they keep Fable on the subscription. (I’m writing this Friday evening and the issue comes out Monday morning - let’s see how this holds up!)

Have a poke around. Just check where your data is going first eh?

To the Task,

Kyle

PS. Full discussion on Youtube. Make sure you are subscribed! 😄 

ChatGPT Work is probably not for you.

If you're reading an AI newsletter, know what Codex is and have spent any time poking around Claude Code…you are already in the loud, weird power-user minority.

And that minority is currently furious with ChatGPT Work.

Work looks like a watered-down Codex. The interface is messy. The names are rubbish. Some options appear on desktop, some on web, some on mobile, and we have all decided this is a personal attack.

All fair!

But…OpenAI is aiming Work at the other one billion ChatGPT users. People who use ChatGPT every day but will never download a terminal, open Codex or learn what an MCP is.

It’s not for YOU as a power user. It’s for the rest of humanity.

The loud minority

Last week I wrote about why ChatGPT Work matters. It gives normal ChatGPT users access to the agentic stuff power users have been hammering for the last six or seven months. That (for me) is pretty cool.

Those of us who are power users got used to AI doing the work very quickly. That’s just what AI does right?

But most people still ask ChatGPT a question, get an answer and then go away to do the actual job themselves. ChatGPT Work is the bridge from asking to delegating.

So yes, Codex is more powerful. It has deeper controls. It can build and run “proper” software. It’s great. I use it daily for faaar too long. If you're already happy there, stay there!

Just know that Work is the easier door for everyone else.

One app, three doors

Admittedly it’s a little confusing now. We have three different ways to use ChatGPT.

The easiest way to choose is to decide what you want back from it.

Use Chat when you want to think something through. Ask questions. Compare options. Research an idea. Argue with it. Get clearer on what you actually want. The chat itself is the output.

Use Work when you already know the outcome and want a finished thing back. A report. A presentation. An analysis. A research pack. Something you can open, check and use. The output will be a file or some sort of artefact.

Use Codex when the finished thing is software. A site, app, database, automation or technical system.

The ten-second version:

• Think: Chat

• Make: Work

• Code: Codex

There is no "best" door. Depends what you need back.

Chat first, then hand it over

And we can chain them!

For example I used Chat to work out the argument for my livestream presentation. We went back and forward until the structure was right. Then I copied the conversation link into Work and told it to build the deck.

Literally shared the chat. Copied the link. Moved to Work mode. Pasted the link and said “let’s work on this”.

Chat helped me decide then Work made the “thing”.

That handoff is useful because giving an agent a fuzzy brief just lets it produce polished rubbish without bothering you for a while. More autonomy does not rescue confused thinking. It merely lets the confusion run for longer without you knowing!! Whilst burning up your usage.

So…start in Chat when you are unsure. Work out the audience, the outcome and what good looks like. Then send the settled thinking to Work to get the “thing” done.

Give Work one proper job

Work can do, well, work. BUT don’t overburden it.

If the task is TOO big then Codex is going to be a better tool.

Instead with Work stick to relatively discrete, simple jobs to start.

For example: pick one job you already understand. Maybe turn a meeting transcript into actions and owners. Compare three supplier quotes. Turn an approved outline and source material into a presentation. Build the weekly report in the same format as last week. These sort of jobs with a defined input and output. Something you can look at and think “yup, it did a good job!”

Give it:

1. The outcome you want.

2. The source material.

3. The boundaries.

4. What "done" looks like.

Your prompting skills still matter because prompting is communication. The Prompting Fundamentals playbook applies even more when the AI can run off for 12 minutes and do a load of work without you. You still NEED to communicate what you actually want if you want good results.

Oh and as always check the result. Sources, numbers, links, the lot. A nice-looking deck can still be full of nonsense. Check the work!

So. If you are a power user your task today is to keep moaning about ChatGPT Work and the new app! 😛 

Everyone else should open it and give it one real job.

To the Task,

Kyle

https://youtu.be/Jo9koewf764 - watch now or save for later

OpenAI and Anthropic are fighting over us.

More accurately, they are fighting over the people hammering Claude Code and Codex all day. But still…we are in the middle and the divorced parents are trying to prove who loves us most.

So we got a pony.

On Sunday Anthropic extended Claude Fable 5 through July 19 and kept Claude Code's weekly limits 50% higher.

Twitter tweet

Exactly 57 minutes and 53 seconds later, OpenAI removed Codex's five-hour restriction and reset everybody's usage.

Twitter tweet

In those 53 minutes I burned 50% of my weekly Codex usage but I knew Tibo was about to drop another reset on us.

I was very happy to be right!

One Day. One Billion Tokens.

One heavy day last month I got through roughly 940 million tokens in Codex. Call it a billion because I was close enough and it makes the maths less annoying…

At Fable's API list price, one billion tokens would cost $10,000 if every token was input or $50,000 if every token was output. The real bill would sit somewhere between those limits depending on the mix and caching - probably $15,000 for the day.

One day mind you.

That is not what I paid. Phew.

I paid for a normal subscription. Like a sane person.

The labs are temporarily converting five-figure compute habits into a two or three-figure bill. They are eating the cost because they want the power users, the developers and the businesses who will eventually spend proper money through the API.

This is a subsidy window.

It is also why people are getting (very) twitchy about Fable disappearing.

Why Fable Keeps Coming Back

Fable is Anthropic's halo model. It gives Claude the claim to having the best model available on a subscription.

The problem is that GPT-5.6 Sol is now extremely close on the Artificial Analysis intelligence index, while costing about a third as much on their benchmark. Grok 4.5 is hanging around the frontier too.

If Anthropic removes Fable from subscriptions, Claude's paid offer suddenly looks a lot less exciting at exactly the point OpenAI is piling on pressure. It could go from 1st position to 3rd position (behind Grok!!!).

So Fable gets another week.

My guess is that Anthropic wants Opus 5 ready around July 20 or 21. That is a guess, based on the extension ending on the 19th. If Opus 5 is not ready, I reckon Fable gets extended again…

Anything but lose position to ChatGPT.

Every Reset, The Builders Go Feral

My WhatsApp group has about 100 people in it. Every time Anthropic or OpenAI resets the meter, people drop what they are doing and pile back into building.

I was working in Codex until about 11pm on Sunday because I expected the reset… adn so were many people in my community. Getting things done whilst the going is good.

Rob Hallam took it much further. Over the last couple weeks he bought four $200 Claude accounts so he could keep Fable running, then stayed up all night trying to finish work before access disappeared. He ended up in hospital with stress and panic-attack-like symptoms.

Twitter tweet

To be clear, the model did not put him in hospital. The stress and binge-work cycle did. Rob said that himself. This is not hustle porn. It is a sign that shifting limits and rubbish communication are making otherwise sensible people behave irrationally. Get well soon Rob!

Rent The Intelligence → Own The Output.

There is a REAL urgency though.

Right now we are getting tens of thousands worth of usage for a hundred bucks a month or less.

The cheap state-of-the-art AI will not be here forever. And when subscriptions stop being subsidise (because, you know, the companies want to make a profit!) we are left with escalating AI bills.

This means we have a window of opportunity to build and create.

Use the rented intelligence while it is cheap. Turn it into things that remain after the reset disappears:

• Code

• Products

• Systems

• Audience

• Revenue

We lease to buy if that makes this clearer. We spend a little now (a few hundred dollars on AI subscriptions) to build assets that can cash flow life changing amounts of money.

There is a darker edge here too. People who turn cheap access into skills, assets and cash now can afford better access later. People who wait may find the tools more expensive at the same time their existing work is under pressure.

That is the K-shaped economy warning- a split between those who use AI to accelerate now and those who ignore it.

Those who grab the opportunity and run with it will be able to make money and thus continue to afford the use of AI.

Those who do not will no longer be able to enter the game because the cost of AI will increase. They’ll be locked out forever.

This is possible outcome, not an economic law carved into stone. But I would rather you start learning now than discover in two years that the useful stuff costs a grand a month.

The answer is not to panic harder. Build something.

Your First Job: Build

Your first project should be a toy. Open Codex, tell it what you want, look at what it gives you and keep talking until it works. You do not need to read the code. You need judgement, feedback and a real problem.

The First Steps in Vibe Coding guide will get you from zero to a working app without the technical faff.

Then build a second thing somebody might actually pay for. Start with the problem and the buyer rather than disappearing for six months to build your grand vision. My AI Business Ideas playbook walks through that bit.

And use the expensive models where they matter. I use Fable for the big hairy planning and judgement, then pass implementation to cheaper models like GPT-5.6 Sol in Codex. Better brains for the hard decisions. Cheaper hands for the build. I wrote a full guide on this recently - how to work with multiple AI tools.

If you are already ahead, bring someone with you. Businesses are desperate for people who can translate this stuff into normal English. I am teaching the full workshop model on Wednesday at 6pm UK time (tonight), and there is a recording if you are watching the football!

The meter will come back. Build before it does.

To the Task,

Kyle

Watch now of save to watch later - https://youtu.be/dypPkH3Kmwc?si=16rBQfUuKgVOB97p

OpenAI released ChatGPT Work yesterday and everyone has the same response:

What the hell is this?

Fair!

My Whatsapp chat is full of peeved people. And…I get it.

OpenAI now has Chat, Work and Codex inside one app. Then you can choose a model. Then you can choose an effort level. Then some options appear on desktop but not mobile, or on mobile but not web, depending on whether the rollout gods like you today.

Twitter tweet

It's a bit of a mess!

BUT…underneath the rubbish naming, something very important just happened. And I think most people missed it.

For 1 BILLION people just ChatGPT stopped being somewhere you ask questions.

One App, Three Doors

OK so what do we actually have now.

The simple version:

• Chat: ask a question and get an answer.

• Work: give it an outcome and let it plan, act and deliver.

• Codex: build software with the full developer controls visible.

Same house. Three doors.

The weird bit is that the walls are a bit flimsy. HOW and WHERE you access these three modes depends on what tool you are using:

• Phone app

• Web interface

• Desktop app

On the Desktop app you switch between Work and Codex. They seem pretty damn similar though - just some visibility differences. You can also pop up a small Chat window for quick questions. But otherwise Chat is gone.

On the web version you can switch between Chat and Work.

On the phone version you have Chat and Work and Remote (to Codex).

Oof. Complex. And annoying to relearn.

That is why power users are moaning. It feels redundant.

But there are millions (nay, 1 billion) people who will happily use ChatGPT every day and will never install Codex, Claude Code or OpenClaw. They hear "coding agent" and do one.

Give them a toggle marked Work though…nice! They’ll get that. And because of that they’ve just been handed agentic AI.

The Unit Of Work Changed

My livestream presentation was made with ChatGPT Work.

I gave it screenshots from Twitter and a long voice note with me blabbering about the launch. It worked for about 12 minutes, spun up three sub-agents, checked official sources, found outside research, outlined the story, made the slides with ImageGen and packaged the lot into a PDF.

I steered it. It did the faff whilst I had my coffee.

If you've already been turning Codex into a personal assistant, that sounds normal. So what? But for a standard ChatGPT user it is a proper change.

Chat has trained us to ask, "How do I make this report?"

Work lets you say, "Make the report. Here are the meeting notes, the source files and the format I need. Bring me back a finished draft."

And that power is now in the hands of everyone.

The Fourth Leap Goes Mainstream

Ethan Mollick talks about four big jumps in modern AI.

GPT-3.5 made the whole thing visible. GPT-4 made it useful for real work. Reasoning models made longer, harder problems possible. Then late last year we got agent systems that could actually go away and complete jobs.

OpenClaw, Claude Cowork and Codex already did this. ChatGPT Work matters because OpenAI has put the same basic behaviour behind a normal-looking toggle in the product everybody already knows.

Distribution matters.

Give It One Real Job

OK how to actually get started with this? This is not about the tech. It’s about your skill level. You need to up your delegation skill.

First up, don't start with a grand plan for an autonomous AI company. Yeesh.

Pick one job you already do that takes about an hour. A report. Meeting notes that need turning into actions. A pile of files that needs analysing. A voice note and screenshots that need becoming a presentation.

Then give Work four things:

1. The outcome you want.

2. The sources it should use.

3. The boundaries it must stay inside.

4. A clear definition of what done looks like.

Your prompting skills still matter because prompting is communication. The better you can explain the result, context and boundaries, the less rubbish you get back. The Prompting Fundamentals playbook still applies. Work just gives those instructions to a system that can keep going for 12 minutes instead of firing back one answer.

Update the app. Find Work. Give it one real job.

To the Task,

Kyle

https://youtu.be/aL2sNhnNNcM - watch now or save for later

Everyone is switching models again!

Fable got a stay of execution until Sunday! ChatGPT just dropped Sol, Terra and Luna! xAI released Grok 4.5!

Hell, even Meta are back in the game with a new model. Whaa?

Nice! More competition. Maximum boost!

We are drowning in state of the art models. Including Fable.

Fun if you still have usage. Less fun if you burned through your Fable allowance before the deadline and then Anthropic wandered back in like, "actually lads, have five more days."

Annoying.

But I reckon the bigger point is not Fable. Or Sol. Or Grok being "Opus-class" (we shall see, Elon, we shall see…).

The wrong question is: which model wins?

The useful question is: how do they work together?

The model is not the system

The model is not the system

Models change. But you can’t be constantly switching or you’ll get NOTHING done.

Instead focus on your workflows. And make the model agnostic.

If your entire AI workflow is "open the smartest model and ask it everything", you are going to have a miserable few years. The smartest model will change - weekly. Or daily if this week is anything to go by! The price will change. The access rules will change.

Change is the only constant as some smart Greek who I can’t be bothered to look up said. ChatGPT would have known…

This week it is Fable, Sol, Grok and (checks notes) Meta Spark 1.1. A year from now it will be other names.

So the durable skill is not brand loyalty. It is workflows that can swap models in and out without collapsing. I wrote about this from the pricing side in the Fable is going away issue. Frontier intelligence is getting rationed. Sometimes by money. Sometimes by policy. Sometimes by capacity.

So the smart play is to use the cleverest model where it actually matters. Use the cheaper thing where it is good enough. Connect the two.

Use expensive intelligence where it matters

Planner model. Worker model. Stop making the expensive one do everything.

The simple split is planner model and worker model. Or an orchestrator and its agents if you fancy.

Planner model: strategy, architecture, judgement, trade-offs, review. This is where you use Fable, Sol or whatever the current expensive biggest brain is.

Worker model: file edits, tests, implementation, cleanup, repeatable tasks. This can be Codex, a cheaper GPT model, a hosted open model, a local model. Whatever clears the bar.

On our live AI Hour call last night Adam talked about how he recently used Fable to scope out a project to convert a PHP project he made 20 years ago into a more modern form. Fable got to work on the strategy and the delegated step by step work to Opus and Sonnet. A big chunky project that took the smaller AIs 4 days(!) but meant saving potentially thousands one tokens by intelligently using the right model for the job. Super valuable skill.

ClaudeDevs had a good version of this on X: use Fable as the advisor, then let a cheaper executor call it when it needs guidance.

Twitter tweet

As I mentioned in the last newsletter I spent about 900 million tokens on Codex on one heavy day (June 14th to be precise). If I ran that through Fable API pricing, the back-of-the-envelope number was around ~$15,000. For one day.

This is why we need to get smarter with how we use different tools.

I personally do not want to remortgage my house because I asked the clever model to fix a typo in a test file.

So do not use the smartest model just because you have access to it. Paradoxically that is NOT a smart thing to do.

But HOW?

Put the tools in the same room

One shared room. One handoff.

The practical bit is much less mystical than people make it sound.

You need a shared folder.

For most of this work, that means a GitHub repo. If that phrase makes you feel itchy, think of it as Google Drive with version history and fewer vibes. I’ve talked about how to set up Github in this Github 101 guide.

Could you use Google Drive? Technically? But I wouldn't.

GitHub is better when multiple agents are making changes because it can track what changed and what needs merging. Google Drive is lovely for documents. It is a bit crap for agent handoffs. Because they’ll all be in there working on the same files and making a mess of it all.

This is the same reason I keep talking about building an AI brain / shared vault. A shared workspace for ALL your AI tools. (Also the topic of our AI Hour live call last night).

Inside the repo you add boring text files:

• README.md

• AGENTS.md

• CLAUDE.md

• tasks.md

• decisions.md

(By the way MD just means Markdown. Basically a stripped-down text file. Less scary than it sounds)

Tell Claude/Fable to write the plan and put the decisions in the repo. Then tell Codex to pick up the work and implement it. That’s (at base) it.

Start manual. Add machinery later.

As well as having a shared workspace we can also directly connect out tools to one another in various ways.

There are three ways to connect the tools.

First: manual handoff. Open both tools from the same repo. Claude writes the plan. Codex does the work. Claude reviews. You are the messenger. That works. It is also a bit of a faff because you become the bottleneck.

Second: plugins. Claude Code can call Codex. Codex can call Claude Code. Both have plugins that you can install in the other. Or if you are using Cursor then install both!

Third: MCP, CLI or custom wiring. This is where the tools talk directly under the hood. This is the most “advanced” but bizarrely probably also the easiest. You can literally ask the AI:

I want you to connect this project to Codex so you can 
hand implementation work off to it. 
Work out the best CLI or MCP route and set it up.

That is it. Seriously. The AI is smart. It’ll work the rest out.

My beginner loop

Start from the best planner. Push the work down. Review before commit.

This is the version I would use if you are starting from scratch today. This WILL change. Might not even be the same a month from now. But the principal holds.

Start in the best planner. For me, right now, that is Fable. I give it the messy brief and tell it to interview me. Then I switch to voice and yap into the microphone.

Fable turns that into a project plan, file structure and shared instruction files.

Then Codex implements. It has a LOT more usage so can work on tasks doggedly for long horizons without bankrupting you.

Then the work goes back up to Fable or another stronger reviewer. It checks the result and sends fixes back down.

That loop itself matters more than whether the model name on the box is Fable, Sol, Terra, Luna, Grok or whatever else gets announced while I am trying to have breakfast….

Do not marry models

The names change. The loop matters.

The warning is simple:

Do not just use the smartest model because you can.

That’s dumb(!).

Instead use the best model for the job at hand.

One plans. One writes. One reviews.

Remember that the. model names will change. The limits will change. The prices will change.

Your job is not to predict which logo wins next week. Leave that to the Twitter bros.

Your job is to build the loop so each release does not matter.

To the Task,

Kyle

https://youtu.be/WRPxqOXpj-Q - watch now or save for later

Fable is gone.

Well shit.

My Whatsapp Group has been abuzz with working out what the hell to use our final moments with Fable on. And then mourning.

Or so I wrote yesterday before we got the news of a reprieve. A stay of execution:

Twitter tweet

This obviously messed up the newsletter a bit. BUT everything below is still accurate. It’s just….delayed by 5 days ha!

As of Tuesday July 12, 2026, Anthropic is moving Claude Fable 5 out of normal paid plans and into usage credits. The API page has it at $10 per million input tokens and $50 per million output tokens.

Now, you might be wondering if that’s a lot. Or even what is a token? How many input and output tokens could a mere mortal use in a day? Let’s run the numbers quickly.

Here for example is me using 900M tokens on Codex in a day (and what a day ‘twas!):

That ONE day, if I had been used that many tokens on Fable would cost me between $9,000 and $45,000

Coding tasks tend towards the input so probably around $12,000-15,000 for the day.

I think my business partner Harms would have been a bit miffed.

Using Fable from now on will be the reserve of companies or the very wealthy.

And I get why people are annoyed. For the last week, a lot of people finally saw what these models can do when they are given actual work.

I used it on finance stuff I had been dodging. I used it on a book project I did not want to outsource to slop. I used it on business systems that had been stuck for months. And Fable just tore through everything.

And now the clever thing is behind a meter. A very expensive meter. Bugger.

Zoom out

The pricing story will get all the shouting and screaming Fair enough. People got a taste of Fable, then Anthropic put it behind usage credits. Cue moaning, mourning, and screenshots of token maths.

But we need to step back and remember the last few years.

A frontier model arrives. Everyone falls in love with it for about five minutes.

Then the next lab catches up, the price drops, the open models trail behind, and what was once science-fiction becomes boring infrastructure.

We get VERY used to our models very quickly.

We will have other models like Fable shortly. Probably this week with ChatGPT Sol. Or if not definitely in the coming months.

The direction is obvious. Fable feels rare and special because we are standing too close to it. In a year, probably less, this level of intelligence will feel normal.

Think about this for a moment:

One day, in the next 2-3 years, we will have a Fable level model running locally on our iPhones.

Read that again and really let it sink in.

Our models are getting better, cheaper and smaller all of the time. ChatGPT 3.5 is just 3.5 years old. And it was TERRIBLE (seriously, go try it on the Playground for a laugh).

It’s taken 3.5 years to come from ChatGPT 3.5 to Fable. What will we have in 3.5 years from now? THAT is what we need to keep our eyes on here.

What happens when its cheap?

When a model is expensive, you ration it.

You use it for the big judgement calls. The ugly personal problem. The business mess where being 20% smarter changes the answer. That was the point of last week's Fable question issue: do not waste the expensive brain on tiny tasks.

But when the price falls, behaviour changes. What happens then?

We will one day have this Fable near-human level intelligence for basically free. On our phones. On our local devices.

That brings us to the darker side.

If you have used Fable over the last week.

And if you can imagine Fable being made free and unlimited.

Now tell me that jobs are safe.

It becomes a very difficult argument to make.

And remember businesses do not even need a model to be free. They need it to be cheaper than the human task it replaces. And when Fable-ish capability gets fast, cheap, and boring like a utility, a lot of white-collar work starts looking wobbly.

The AI doesn’t have to take entire jobs. No need. Remember that jobs are bundles of tasks. If AI takes 30%, 40%, 50% of those tasks, the job title might survive while the income, headcount, and employee bargaining power get quietly chewed up underneath it.

This is where I get annoyed at the "it'll all be fine" crowd. They accuse people like me of being a Doomer. No - I’m a realist. And after seeing what Fable can do I’m doubling down on that. Our comfy 9-5s are not safe.

Build something you control

You cannot rely on your company to handle this well. If you are employed you need options.

Maybe your employer has a clear AI strategy. Maybe they have tooling, training, governance, a proper plan, the lot.

Maybe.

Or maybe they faff around for two years, buy a licence, form a committee, and then panic-cut costs when a competitor uses AI faster than they do.

If you work inside that machine, you do not control the machine. Sorry.

Or shit maybe they do know what they are doing and you still aren’t safe. Because the first moment they can drop you and replace you with AI they will. Corporations want our loyalty but the honour does not extend in both directions. They will train the AI on your daily work (screen recordings, emails, meeting notes, any and everything they can get their hands on) and then switch you out for an agent that does the task for 100x lower cost.

And maybe even better than you…

So what do we do?

So build something you do control. A side income stream. A consulting offer. A tiny first product. A niche audience. A service. Something you build yourself and control.

I’ve actually built a whole 10 week course on building a business with AI. Totally free, no need to even register.

Is it easy? Nah, not easy. But increasingly we don’t have much of a choice.

We either wait for the sword for fall. And then complain that we didn’t see it coming (we did).

Or we take responsibility and get on with building something that we control.

If you are looking for a starting point I say lean into the chaos! Let the confusion and fear of AI be what drives your business. Instead of being driven by it. This is why I keep banging on about workshops and advisory work. EVERY business on earth right now knows they need help with AI.

And they have no idea what to do about it.

You do. You are in a very very small % of people who are paying attention to AI. I know this because you are here reading a debrief about Fable. That’s pretty niche info you know.

If you can explain this stuff clearly and turn chaos into action, that is valuable to businesses. Extremely valuable. Several thousands pounds/dollars an hour valuable.

I’ve got 250+ students are building around that model now. Workshops, consulting, implementation, advisory. If you want to know more I’m doing a live webinar tonight at 6PM London time: aiwithkyle.com/webinar.

To the Task,

Kyle

Watch now or save to watch later: https://youtu.be/P4-N2VQIWiI

Fable is back. But not for long.

Anthropic has restored access to Claude Fable 5, and for a few days the expensive thing is sitting inside normal paid plans for up to 50% of weekly usage limits. That window closes on July 7, 2026. 

After that, it moves to usage credits and the API price is a cool $10 per million input tokens and $50 per million output tokens. If you don’t use the API and have no idea what it means I’ll translate: effing expensive!

So yes, this is time-sensitive. It closes

This is basically Oracle of Delphi energy. You get to walk up to the very clever machine and ask it the big question…

And most people are still going to ask it to tidy an email. Please don't! Hell if you are going to do this sell me your usage instead ha!

Do the 20% test

OK so what should you ACTUALLY use your (limited) Fable on?

This is the filter I would use: Ask Fable things where being 20% smarter changes the answer entirely.

Is it actually 20% smarter than all the other models? Maybe. Doesn’t matter. Use this as a mental framework. Basically think of it as the smartest person you’ve ever talked to…what would you ask them?

Judgment calls. Tricky trade-offs. Untangling messy situations. The question behind the question. The hard stuff!

Do not use it for boilerplate code. Do not use it for formatting. Do not use it to summarise something Opus or GPT-5.5 can already summarise perfectly well. That is using a freight train to pick up your shopping.

Total waste.

If you have a messy business decision, give it that. If you have a huge project repo and you cannot see the shape of the thing anymore, give it that. If you have a personal or professional decision you have been dodging because the answer is probably inconvenient, give it that.

Build the question first

I’m seeing a lot of people absolutely paralysed about what to ask Fable. With such limited usage and such a tight time constraint I get it. Easy to get blocked.

Here’s a concrete tip…use a “dumber” model to help you construct your question.

Do not sit there staring at the Fable box trying to write the perfect prompt.

Use a cheaper, slightly dumber model as the workshop. (God, it feels weird calling Opus 4.8 or GPT-5.5 dumb… but you know what I mean.)

Brain-dump your situation into Opus or GPT-5.5 first. Tell it you have one chance to ask a much smarter model a question. Get it to interview you. Get it to pull the uncomfortable context out of you. Get it to draft the full one-shot prompt.

Then pressure-test it:

• what is the prompt missing?

• what would Fable need to know if I get no follow-ups?

• what assumptions am I hiding?

And because your usage may be capped it’s important to ask in one go. You don’t want to end up answering lots of back and forth with Fable as each turn will eliminate tokens. So get it right first time as much as you can!

Assume you get one prompt per day. Especially on the $20/month plan!

Maybe a handful more if Claude is feeling generous. Maybe five if you are on the $200 plan and burn through your allowance like an idiot. On $200/month I’m getting 4-5 every 5 hours. And maxing them out every time!

So the prompt needs three bits:

• context: who you are, the situation, what you have tried

• question: one specific, high-stakes thing

• safety net: assumptions, counter-arguments, likely missing context, and what it should do if it would normally ask a follow-up

This is different from normal chat prompting. Normal prompting is often a conversation - we have “turns” - going back and forth. With Fable it’s much closer to briefing an expert before they walk into a boardroom and you are not allowed to speak again.

Ask the nasty thing

The best Fable question is probably hiding behind something uncomfortable.

What decision have I been avoiding?

Where am I working hard on the wrong thing?

What would embarrass me if someone brilliant looked at my business for an hour?

I gave Fable my AI brain and asked versions of that. It has my project logs, my decisions, the stuff I have been working on, the stuff I said I would do and then quietly wandered away from. If you want to build that sort of shared context, the AI brain guide is here.

And it called me out. Brutally. It was quite upsetting!

It basically said I keep killing winners. I start a project, prove people want it, get bored when it starts working, then wander off to the next shiny thing…

Rude. And 1000% accurate. Which is annoying.

That is what this model is for. Not "rewrite this LinkedIn post in a more professional tone." Nah, use it for BIG gnarly questions and problems.

The way I like to think about it is that answers are getting cheap. Good questions are scarce. And that is on you.

One final point for those who may be feeling overwhelmed by having to get full usage out of Fable right now. i) It will eventually come to our normal subscriptions which is cool but ii) more importantly: ALL models will eventually be at this level. It may be 6 months from now. It may be 12 months from now. But it will happen.

So if you feel you’ve missed the boat this time don’t worry. This level of capability will become the norm. Which is both terribly exciting and terribly scary!

To the Task,

Kyle

https://youtu.be/L9MV98e93ow - watch now or save for later

All this fuss about Open source AI is a price war.

Sorry open-source lads and lassies. I know there are philosophical bits. Safety bits. Licence bits. Community bits. All true, all amazing stuff.

But the reason this suddenly matters to normal people is money.

Anthropic and OpenAI need the world to believe that the best AI has to come through expensive, “safe”, American API pipes. Controlled by, well, them.

Meanwhile, Chinese labs are shipping opensource models that are more than good enough for a lot of real work and a hell of a lot cheaper.

The very existence of these opensource alternative undercuts the whole US/closed source mission. And threatens to bring the entire economy of the West crashing down.

So…this is not just about software development principals!

WTF is opensource?

“Open” is doing a lot of work here!

What actually is opensource? The annoying thing is that everyone uses "open source AI" to mean different things.

Sometimes they mean free. Sometimes downloadable. Sometimes local. Sometimes the training data is visible. Sometimes the licence lets you use it commercially. Sometimes they mean "I saw DeepSeek on Twitter and now I am pretending to understand geopolitics."

This is further confused by the fact that most opensource LLMs aren’t actually opensource. They are open-weights. Grr…

The useful split is this:

• open source means you can see the code, the weights, the training pipeline and sometimes even the underlying data.

• open weights means you can download the trained model files

Most of the models people call open source are really open weights. You can get the trained numbers. You can run them somewhere. You might be able to fine-tune them. But you usually cannot rebuild the full thing from scratch because you do not have the training data, process, and boring details behind the model.

Here’s a useful analogy from Dentro.

Closed source models like ChatGPT and Claude are like a taxi. You can tell the driver where to go, but you do not own the vehicle. You definitely can’t take it home and start modifying it - the cabbie would get very annoyed I imagine. And, like black cabs, they are expensive!

Open weights are more like owning a car. You have the thing. You can drive it where you want. You can modify bits of it. You are responsible for the faff that comes with ownership. Most of the time this works out cheaper than taking a taxi but on the whole it’s a little more of a faff.

True open source is the kit car version. You get the parts, the plans and enough information to rebuild the thing yourself. Very few models are truly open-source, indeed none of the well-known production grade ones.

Also whether a model is free to use and whether it’s local are entirely separate issues from the model being open weights.

GLM 5.2 for example is technically open-weights. You can go right now and download the whole thing, all 1.5TB of it. And then run it on your local computer for “free” (minus the cost of electricity of course).

BUT…your local computer is going to have to be a data centre… you aren’t running this on any commercially available computer.

Most local models are opensource. But not all opensource models can be run locally! So don’t conflate the two.

Where does all this live?

Let’s actually have a look at some open-source (open-weights, see even I’m using the term wrong..) models. How do you find them? Where can you download them? What can you do with them?

The first site to look at is Artificial Analysis. Their open-source model page compares open-weight models by intelligence, openness, size and more.

These are pretty much ALL Chinese models and the first non Chinese model is Nemotron (Nvidia).

Let’s have a look at the intelligence against cost-per-task. Basically your bang for your buck:

https://artificialanalysis.ai/?cost=intelligence-vs-cost-per-task

The smartest models are (duh) Anthropic’s Fable, Opus, Sonnet and GPT5.5. They are also the most expensive.

GLM5.2(max) is sitting at a very similar intelligence level (same as Sonnet 5 which came out this week) but at a fraction of the cost.

Twitter tweet

The x-axes here is log scale by the way so the costs are getting much higher in that riht hand side - look how it jumps from $1 to $2 to $3 in smaller steps.

Opensource models are coming in either i) slightly worse but much cheaper or sometimes ii) equivalent quality but much cheaper. That is where the price war becomes obvious.

You do not always need the smartest model. You need the cheapest model that clears the bar for the job. Customer service reply? Internal summary? Classification? Simple data clean-up? A frontier model is often total overkill.

We have a pigeon that keeps flying into the garden and making a mess right now. The sensible method to get rid of it is to wave and clap until it buggers off. The Claude Fable method would be to hit it with an Intercontinental Ballistic Missile.

Often we don’t NEED so much power. And when that power comes at 50-100x the cost it’s silly to use it.

The second site to know about is Hugging Face. This is where the models live. Here for example is GLM-5.2:

The tags at the top tell us what it can do (text gen), its languages, its license etc.

On the right we get the model size - here it is 753 billion parameters (big!).

Hugging Face has almost 3 million other models for you to browse, learn about and download.

But remember that most won’t run on your local computer. Opensource is NOT the same as local! This GLM-5.2 model is the opposite of a casual laptop download - it’ll need a small data centre to run it!

Still incredibly important though. You can inspect it. You can see the files. You can see how this stuff is actually distributed. It’s all right there for you to explore.

China is playing a blinder

This is where the geopolitics come in.

The US labs have a very expensive story to sell. More chips. More data centres. More money. More closed models. More funding rounds. More valuation nonsense.

Scaling has worked so far. Just throw MORE at the problem and the AIs get better. It’s actually pretty amazing. And noone is quite sure why it works…

OpenAI and Anthropic are trying to justify obscene numbers and the market is nodding along because AI is the big growth story. About 2% of US GDP is being invested into AI this year and somewhere north of 50-60% of all VC money is going into AI startups.

It’s a massively intense concentration of investment. In one basket. A basket that is currently not laying eggs? I’ve lost the metaphor…you know what I mean.

A lot is now riding on 2 companies. Anthropic and OpenAI. And their upcoming IPOs (going public on the stock market). Both of their IPOs are based on valuations just shy of $1T with revenues being a tiny multiple of that - much lower than the revenue multiples we expect.

It’s one hell of a big bet. And making such a big bet requires investor confidence. If they get wind of this not working out, of their money not coming back to them, of potentially losing everything then they’ll panic and withdraw.

The IPOs of Anthropic and OpenAI are big litmus tests here. Do the markets have the balls to support the dream? Big, scary risk.

…Enter China.

China is bringing opensource models to market. Not with one model. With a flood of them. DeepSeek, Qwen, Kimi, GLM, MiniMax, Hunyuan and the lot. Some are brilliant. Some are a bit crap, much like US models! But the direction is clear: cheaper, good-enough intelligence keeps arriving.

China's open-weight model crowd is not exactly small.

And if the best closed model is only needed for the hardest 5-10% of work, then a lot of the Western AI economy starts looking very different.

Why would I pay Anthropic or OpenAI a sizeable chunk of my revenue every single month for tokens when I can deploy a Chinese model locally within my organisation, fine tune it and run it basically for the cost of electricity thereafter? Shit, I even get better security because my information is not being sent up and down to Anthropic / OpenAI. It sits locally in a secure server in my office.

The very existence of these cheap, efficient, Chinese local models is a threat to the narrative the closed-source American labs are pushing. It undercuts their whole business model.

Even if very few companies actually deploy local Chinese based models the fact that they even exist as an alternative sheds doubt on OpenAI and Anthropic.

This is why the safety argument from closed labs always feels a bit fishy to me. To be clear: the safety argument is real. You cannot release powerful weights into the world and then recall them. Once a model is out, it is out - Pandora’s box is open. Governments should care about that. Cyber, fraud, bio, state actors…lots of scary stuff.

BUT...

The business incentive is also real.

Closed labs fear open models because they are dangerous. Sure. They also fear them because they are becoming good enough. Good enough kills margins. Good enough gives enterprise buyers wiggle room. Good enough means developers can switch vendors. Good enough means "pay us whatever we ask" stops working.

That is a wedge. If a huge chunk of US market confidence is tied up in AI capex, and Chinese open-weight models keep making intelligence cheaper, this starts to look much bigger than a nerd fight.

If the Chinese opensource models can destablise the Anthropic and OpenAI IPOs and cause sufficient investor doubt then the whole AI industry could come tumbling down, bringing with it the American and subsequently Western economy.

This is not a fight over closed-source vs. open-sourced software development.

This is a fight for who owns the top spot in the global economy: the US or China.

Pick a stack, not a religion

So what do you actually DO?

Do not become an open-source purist. Also do not become a closed-model fanboi. Both are annoying. And you know it.

Instead play it smart.

Use frontier models for genuinely hard work. Strategy. Proper reasoning. Important writing. Research. Stuff where being 5% better actually matters. Use the VERY best you can afford here and keep updating as new models and tools release.

Use hosted models for cheap volume. Either ChatGPT/Claude subscriptions OR open-source subscriptions for more usage. Or both! This is where routine automations, internal workflows live. Anywhere "pretty good and much cheaper" beats "best in class and wildly expensive". This is the workhorse.

Then use local models for privacy, fallback, learning and simple repetitive work. LM Studio is the friendly route. Ollama is more developer-ish but useful. Gemma is worth watching. Chinese models are worth testing if your data/privacy situation allows it. These models will live on your device and deal with the day to day and anything they can’t deal with gets bounced up the chain.

This setup gives you and your business the most flexibility to deal with whatever is coming down the tracks.

To the Task,

Kyle

https://youtu.be/vtS7i66fKFw - watch now or save for later

From the livestream.

The argument in one picture. Because apparently that is what it takes now.

ChatGPT's best new model is here. You can't use it. Sorry!

And no, before everyone starts shouting, GPT-5.6 has not been banned. That's the hype version going around on social media. And as always it’s not quite right.

Instead it’s actually MORE worrying. OpenAI launched it as a limited preview at the US government's request, and normal ChatGPT users are outside the preview gates. We’re locked out.

This is the new precedent.

Gulp.

I said on the livestream this might be one of the darkest weeks in AI. I stand by that. Not because AI is over. The opposite. Because the best AI is shifting from app-store product to permissioned infrastructure. It’s getting TOO good.

Ignore the ban hype

OpenAI announced three models: Sol, Terra and Luna. Lovely celestial branding. Sun, Earth, Moon. Very tasteful…

Twitter tweet

Totally useless for us though.

Because we don’t have access. OpenAI says it plans broader access later, but for now GPT-5.6 is a limited preview for a small number of trusted partners in Codex and the API.

We don’t know the list or how to (har har) get on it.

Sam's follow-up is here too:

Twitter tweet

Annoying. BUT at least Sam is communicating and acknowledges that this is a messy situation. Dario Amodei and Anthropic are not communicating with their users.

Two weeks. Same pattern.

On June 12, Anthropic disabled Fable 5 and Mythos 5 after a US export-control directive. On June 26, OpenAI shipped GPT-5.6, but gated access before normal users could touch it.

Two weeks. Two frontier-models barred from general access.

Fable/Mythos was launched and quickly pulled. Mythos is apparently coming back for approved entities, not for individuals. GPT-5.6 launched in a restricted preview. Slightly different mechanism but same basic result. The average Joe is out of the loop. Sorry Joe.

This is no longer just "can I afford the good model?" It is: are you on the list? Reminds me a bit of certain Rolex watches and Hermes birkin bags - it doesn’t matter how rich you are. That’s not how you get access. This is not a model I’m happy to see spread!

There is a real reason

The government is not just doing this for a laugh. Or because they are idiots. And not just because Dario Amodei has been stirring up fear.

No, there are genuine very real concerns.

These models are getting scary-good at cybersecurity work, long-horizon agent tasks and scientific discovery. All good stuff.

But we also don't need to pretend this is normal software anymore.

All the same mechanisms for good work can be used for bad work. If it can do one it can do the other.

If a model can help discover vulnerabilities, assemble exploit pieces, run agents more efficiently and compress difficult technical work then governments are going to care. They should care. Banks, infrastructure, health systems and government departments are mostly held together with Excel sheets, procurement forms and vibes.

And even with “dumbed” models bad actors were having an absolute field day.

No model is immune to be jailbroken. And to believe this is possible (as the White House seem to be demanding…) is naïve.

All models will get broken open to be used for naughty deeds. That’s part and parcel of this I’m afraid. Up until now this has been a nuisance but these (much) smarter models make this a critical risk.

Build a model-access hedge

OK let’s come back to what do YOU do again. Always we need to come back to this.

This is not a "delete ChatGPT and live in a bunker" issue. (Although if you do have space in your bunker for one more I’d love to know…)

Use the best models when you can. Obviously. They are absurdly useful. But don't build a your life or business around a single-model.

Have ChatGPT as one option, not your only option. Keep Claude, Gemini or another hosted provider warm. Have a poke around with Chinese-hosted models like GLM through Z.ai if you are comfortable with the trade-offs. Start testing local Gemma-style or Chinese models for lower-risk work you want to fully control. I wrote about first steps with local AI models a few days ago.

Also build out an AI brain. Get everything in one place. Your prompts. Your specs. Your source files. Your evaluation examples. Your workflow notes. Your project context. Do not trap the brain of your business inside one rented interface.

Get it into a Github repo that you can move BETWEEN any AI. This means you aren’t screwed when one particular model is taken from us. I wrote extensively about this last week https://aiwithkyle.com/ai-news/build-one-shared-ai-vault 

Basically you are building in your own flexibility NOW before you need it.

To the Task,

Kyle

https://youtu.be/LqkglPK9SUE - watch now or save for later

Your AI work is everywhere. All over the shop.

ChatGPT has one bit. Claude has another. Codex has half a project. Your phone has the voice note. GitHub has the code.

Sure you are doing a lot. Very productive. Lovely.

But the more you create the more mess you make.

The problem is that none of these things automatically know what the other ones are doing. So you become the USB stick. You are the one copying context from one model to another, re-explaining your business, re-explaining the client, re-explaining the project, re-explaining the weird decision you made last Tuesday because Future You apparently enjoys admin.

That’s precisely the opposite of what we want to use AI for. We are just making MORE work for ourselves.

This is why people are suddenly talking about AI brains, Obsidian vaults and personal operating systems. It looks new. It is sorta new. But (as always) the genuinely useful bit is not the flashy viral stuff.

It is mostly folders and markdown files.

Ho hum.

Obsidian is not the brain

You’ve probably seen posts about Obsidian. And think I’m going to focus on Obsidian.

Obsidian is cool. I like Obsidian.

But Obsidian is the HUMAN interface. It is the pretty layer. It gives you the graph view, backlinks, search, daily notes, plugins and all the nerdy little goodies that make a folder of text files feel like a system.

For you, that is useful.

For the AI? Less so. They don’t NEED this.

Claude Code, Codex and the rest do not need a lovely purple graph of your knowledge. They need the actual files. They need the README. They need the decisions log. They need the project notes. They need the bit where you wrote "my god do not touch this payment flow because it breaks checkout!!1!" three months ago and then forgot about it.

That is the brain.

And it’s basically a folder, some sub folders and some text files. That’s it.

Boring bit that gets things done

The base layer here is a markdown file. .md.

Nothing to do with doctors. No stethoscope required.

Markdown is basically a plain text file with a bit of simple formatting. Humans can read it. AIs can read it. GitHub can track it. Obsidian can display it. Claude Code and Codex can scan it before doing work.

Markdown is not sexy. It is aggressively unsexy. It is a Word document with the formatting fluff removed and, frankly, thank you very much.

But boring is useful here. You do not want the memory of your business trapped inside one app's proprietary format. Or heavy files like PDFs and Word docs.

You want files that can be copied, searched, versioned, backed up and read by whatever AI tool you use next month when this week's favourite app inevitably changes its pricing or does one.

The useful files are simple:

• README.md for what the brain is

• AGENTS.md for how AI tools should behave

• CLAUDE.md if you want Claude-specific instructions (hint, point it at AGENTS.md instead so you have one set of instructions)

• tasks.md for open loops and todos

• decisions.md for things that should not be re-debated

• project READMEs for active work

• agent logs for what AI tools actually did

You can of course add more. But generally something simple like this is all you need.

Also don’t worry about how to make all this - AI will do it for us shortly.

Access anywhere

Where should this folder and set of documents live?

We will connect ALL our AI tools to the same set of folders.

So: we need it accessible anywhere. In the Cloud.

Someone asked on the stream whether this could live in Google Drive.

Technically, yes.

I would not.

If you are using AI tools properly, you will soon have multiple agents working on different bits of your life or business. Codex doing one thing. Claude Code doing another. Maybe one thread sorting your newsletter, another checking your website, another helping with a client project.

If those tools all bash away inside the same Google Doc, it gets messy fast. Conflicts. Random edits. "Who changed this?" energy. Lots of faff.

GitHub is built for this kind of thing.

I’ve done another full guide on Github 101 previously so go check that if have no idea what Github is.

Simply put though it’s for online file storage in the cloud. It tracks changes. It lets tools work in branches or worktrees. It gives you a history of decisions. It means you can see that on June 6th you changed the website, added a rule, moved a file, broke something, fixed it, then told Future You never to do that again.

So: use Github here to make your life much easier. Bonus - it’s free.

Do not copy someone else's brain

This is the bit people will ignore - I know some of you will still go and find some guru’s “perfect AI brain” template. Maybe even buy it! Don’t!

You should not start with someone else's template.

I know. Annoying. Templates are comfy. They make the work feel done before it is done. You download a folder structure, put it in Obsidian, stare at the cool looking knowledge graph and think: "Ah yes, productivity."

Nope!

Your vault needs to map to YOUR actual life. Your roles. Your projects. Your clients. Your products. Your weird repeated context. Your privacy boundaries. Your working style.

Your mess. Your very own chaos!

That is why the starting point should not be a template. It has to start with you.

The starting point therefore should be an interview.

Let the AI ask you questions. One at a time. Then answer with your microphone. Yap for 20 minutes. Tell it about the projects, the priorities, the stuff you keep forgetting, the workflows you repeat, the things you definitely do not want it to touch, and the boring context you are sick of retyping.

I personally did this over DAYS when setting my systems up. I kept remembering important things and yapping some more!

The AI can then propose the folder structure for you. It will make up the files, the project folders, the whole shebang for you.

Much better.

Steal this prompt

Use Codex or Claude Code for this. ChatGPT can help design the structure, but it cannot reliably create the folder tree and files on your machine unless you connect it to tools. Codex is free if need to just getting started

Here is the basic prompt:

I want to create a vendor-agnostic AI vault that acts as a shared brain across my tools.

Use markdown files and folders.

Interview me one question at a time about:
- my daily life and roles
- my work and active projects
- repeated context I keep explaining to AI
- workflows I want AI to help with
- privacy boundaries
- how I like AI tools to behave

After the interview, propose:
- a vault folder structure
- README.md
- AGENTS.md
- project README templates
- tasks.md
- decisions.md
- an agent-log format
- a migration plan

Ask me the first question now.

This will live in Github. 
If I do not have this set up help me.

Then talk.

Seriously: actual talk. Use the microphone. Do not sit there trying to write the perfect answer like a LinkedIn thought leader. Ick.

Just talk. Yap. Unload everything.

The AI will sort the structure out afterwards. That is the whole point. We are not doing artisanal folder design here. Let the AI do the work.

The rule: explain it twice, save it

One issue right now - it’s hard to capture “normal” chats into your Github folders and files. Everyday chatbot Claude and ChatGPT don’t do this by default.

Instead use Codex or Clade Code to make sure everything is captured.

Does this mean you need to always use Claude Code or Codex? Nope!

The daily habit is simple: Anything you explain to AI twice belongs in the vault.

For any one off usage of AI the normal chatbot is entirely fine. Those chats are disposable. For anything you want to capture use your new system.

And the magic thing here is that over time your system gets better. The more context and information it has about you the better your AI will be able to help you.

Not on day one. Day one is mostly a folder and a mild headache as you get it all set up. Sorry…

But after a few weeks, your AI tools stop starting from zero. They stop asking the same questions. They stop making the same bad assumptions. They can read the vault first, understand the context, then do the work.

If your coding tool knows your business it’ll stop making errors that seem “dumb” to you. “What the hell are you doing Codex you KNOW I don’t sell that those services?” sort of annoyances disappear because your AI tools will start to collate info from across different domains of your life.

This is also why chat history is not enough. It doesn’t pass between tools. We need a centralised location for that!

Obsidian is still useful

Full circle!

Should you use Obsidian?

Yeah, probably!

It’ll give you a great human interface to “see” your system.

It is free for the basic local version. It is good. It makes markdown folders easier to browse. It gives you backlinks and search and graph views and plugins and all the nice human bits.

Do you need to pay? Probably not! The paid plan is for synching. But we are using GitHub for that (for free) so we’re good!

Set the GitHub repo up locally, open that folder as your Obsidian vault, and let GitHub Desktop or your AI tool handle the pushing and pulling. Tada! Saved some money.

Again, keep the layers clear:

• folder = the brain

• markdown = the memory

• GitHub = history and sync

• Obsidian = human interface

• AI tools = workers

Grossly simplified. Also useful.

Start small

You do not need a perfect personal operating system by Friday.

Please do not spend three days building a 97-folder cathedral to productivity and then never use it. Productivity porn is appealing. But valueless.

Start with one repo. One README with a basic description. One AGENTS file with simple rules for your AI tools. One decisions file. One task list. One project folder.

Then USE it.

Tell every AI tool to read the vault first. By default they’ll always check AGENTS.md first (that’s its purpose) so put important instructions in there. Again the AI will help you set this up so don’t worry about the how.

Have your AI update the vault when a decision is made. Have it write logs when it changes something. Have it add repeated context instead of letting it disappear into a chat thread you will never find again.

And over time it’ll grow and improve. For now though: keep it simple.

To the Task,

Kyle

https://youtu.be/fjUARfuLsis - watch now or save for later

Fable came back from the future and got immediately confiscated.

That sounds dramatic. It is.

For about four hours I had access to the model everyone is now mourning. People were using it to debug apps, inspect businesses, plan their lives, rethink whole projects, and generally do the kind of meaty work that normally takes days of faff.

Then it vanished…

The livestream story will probably change (again!) by the time this lands. Maybe Anthropic gets Fable back online. Maybe the US government carves allies back in. Maybe they relabel the whole thing Claude Opus 4.9 and pretend everything is fine…

But the important part is not "will we get our shiny new model back?"

The useful part is this: frontier AI just stopped being normal software and started looking like strategic infrastructure.

Fable was different

Same family. Different access model.

Quick recap on what Fable actually is…

Anthropic launched Claude Fable 5 and Claude Mythos 5 on June 9. Fable was the public commercial release. Mythos was the more permissive, trusted-access version for cyberdefenders and infrastructure partners.

Anthropic described Fable as a “Mythos-class model” made safe for general use. And it is (was?) very very good.

I think the reason people have been weirdly emotional about losing it is that Fable did not feel like a faster chatbot. Ethan Mollick wrote about the shift from doing to commissioning. Simon Willison called it relentlessly proactive. Matt Shumer was saying the productivity gap felt huge.

That matches my tiny, annoying glimpse of it. And others I’ve talked to.

It felt (dare I say it…) an awful lot like AGI…

Shh shh shh we don’t say that around here!

The timeline is already a mess

And then it got banned! Ruh-roh.

The timeline is a mess and we’re still getting details. But let’s try to break it down.

Fable launched. People fell in love with it. Then the whole thing got dragged into export controls.

Twitter tweet

Anthropic's June 12 statement says the US government issued an export-control directive suspending access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign-national Anthropic employees.

Foreign-national employees.

That means if you are a non-US citizen working inside an American AI lab (like, I dunno, Andrej Karpathy!), building the actual model, the order says you cannot use the model.

Anthropic says the directive arrived at 5:21pm ET, did not provide specific details, and referred to a possible jailbreak. Their position is basically: yes, there was a narrow bypass, but the demonstrated capability is available in other public models too, and no universal jailbreak has been shown.

Apparently Anthropic had 90 minutes to comply.

The government-side story is different. David Sacks' version is that a credible partner (we know believe this to be Amazon…) found a serious guardrail issue, Anthropic was asked to fix it or pull the model, Dario refused, and the export control followed.

Worth reading in full:

Twitter tweet

So we have two stories:

• Anthropic says that the jailbreak threat was vague, rushed, and not grounded in enough technical detail.

• The government says Anthropic talked up a cyber weapon, got warned about the safety layer, then downplayed the problem when it became inconvenient.

Both can be partly true. And probably are.

I do think Dario has slightly FAFO’d here. If you spend months telling governments your model is scary, dangerous, and needs regulation, you cannot be too shocked when a government says: "OK then. Regulated."

But it is also problematic the US government can force Anthropic to yank a model globally based on a process nobody can see, a technical standard nobody can inspect, and a nationality rule nobody can enforce without turning every AI login into border control.

The nationality rule

This is the bit that turns a model pause into an access-control story.

The model being paused is annoying. But it is what it is.

What is much more important is the introduction of a foreign-national rule. That’s a precedent.

Because how do you enforce that? IP address? Lol. VPN. Billing country? Nope. Company account? Maybe. Passport scan? Driving licence? Face check? Some sort of live "prove you are the right citizen" identity layer?

From a practical POV Anthropic couldn’t “screen out” non-US citizens in 90 minutes. So they shut everything down - it was the only safe play.

Very likely though we are moving into a new world of AI control. AI companies collecting our details and tying them to our accounts. This is a whole new world. You get know-your-customer checks. You get passports. You get government-approved access tiers. You get a centralised record of who is allowed to use which model for which kind of work.

Maybe that is necessary at the frontier. Maybe the cyber and bio risks really do demand it. I'm not doing the silly libertarian thing where every rule is tyranny and every safety team should do one.

But there is a trade-off here.

If frontier AI requires identity-gated access, then AI is not becoming democratised in the way people promised. It is becoming licensed infrastructure.

Who gets the licence? Who gets cut off?

Who gets the good stuff and who gets the leftovers?

We don’t know the answers yet but let’s hazard a guess. These rules will be laid down by the White House and billion dollar corporations.

It’ll be for the rich, big business and approved nations.

Closed access is not ownership

This is a good reminder: if you are using a subscription or an API, you do not own the capability. You have permission to access it until someone else changes their mind.

That someone might be Anthropic. It might be a cloud provider. It might be a regulator. It might be the US government at 5:21pm on a Friday.

That is not a reason to stop using frontier models. That would be mad. These things are still the most useful tools we have.

But it is a reason to stop depending on one model, one lab, one country, and one access layer.

At minimum:

Know which layer you control.

Use fallback models. Keep a second provider ready. Know which workflows can move from Claude to ChatGPT to Gemini to a local model without collapsing. Keep exports of your important prompts, specs, source files, and context. Do not trap the brain of your business inside one rented interface.

And start having a poke around with local models. I’ve written a guide on Local LLMs 101 here. Not because local models are better. They generally are not…. Not because open source magically saves you. It does not. Most people do not have the hardware, the patience, or the cojones to run the really big stuff well.

But local gives you a layer you control.

Use LM Studio. Download something from Hugging Face. Run a small model on your laptop. Rent a GPU server for a weekend if you want to get spicy. Learn what it can and cannot do.

Do this as a learning experience if nothing else.

The scarce skill is bigger tasks

The irony of all this is that the practical skill did not disappear with Fable.

The model is gone for now sure. But the directionality is in play.

Nate B. Jones had the useful framing: the new scarce skill is task imagination. Coming up with work big enough to hand to an agent that can run for hours or days.

The models will keep getting better. Maybe Fable comes back. Maybe GPT-5.6 does something similar. Maybe Claude 5 is Mythos with a new hat. Maybe some compound model from OpenRouter gets you close enough by stitching systems together.

Any which way we are moving towards a Fable-shaped world.

The moral, because yes, apparently Fable needed one.

There are still open questions.

Does Fable come back quickly or does this become the precedent? Will governments review every Mythos-level release? Do “allies” get carved back in or does access stay nationality-based? Who decides what counts as a serious jailbreak? What happens when Chinese labs keep pushing while American labs get tangled in their own rules?

I don't know. Nobody does.

But I know what I would do this week:

Do not depend on one model. Build fallbacks. Learn local models enough that they stop seeming like arcane nonsense. Keep your important work portable. Stop thinking in tiny prompts. Start thinking in BIG projects.

To the Task,

Kyle