Fable got confiscated

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

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.

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:

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

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:

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.

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