The Fable 5 Shutdown

Why the Government Shut Down the Most Powerful AI Ever Released

Fable 5 proved it could do the work. Then it proved nobody could afford it. Then the government proved it could be switched off.

Photo by Nicholas Cappello on Unsplash

Hundreds of millions of people were using it on a Tuesday. By Friday night, they couldn’t touch it.

You probably heard about the shutdown. The US government ordering Anthropic to cut off the model for foreign nationals worldwide. A first for commercial AI. The shutdown made headlines. The walls that preceded it tell the real story.

In seven days, this model hit three walls. Each one redefined what impossible means.

Wall One: Capability

Fable 5 launched on June 9. Within hours, the benchmarks were staggering.

Stripe handed it a 50-million-line codebase and asked for a migration. Fable 5 did it in a single day. Anthropic says that would have taken a team over two months by hand. It scored 93% on Hex’s analytical benchmark, saturating their standard evals so thoroughly they built a new one. It taught itself to beat Pokemon FireRed from raw screenshots alone. No maps. No game state. Just pixels and forty hours.

The base model behind Fable is called Mythos. In April, Anthropic built Mythos and decided it was too dangerous to release. Instead, they gave it to a short list of organizations — Amazon, Google, Microsoft, JP Morgan, power companies, water systems, hospital operators. In about a month, Mythos found more than 10,000 high- or critical-severity security flaws. One had been hiding in OpenBSD for 27 years.

Fable 5 is that same engine with safety guardrails. The longer and more complex the task, the bigger its lead gets. On SWE-Bench Pro, it scored 80.3% against GPT-5.5’s 58.6%.

The first question was simple: can the AI do it? Answer: yes.

Wall Two: Cost

Capability wasn’t the bottleneck anymore. Price was.

Fable 5’s per-token pricing runs roughly $10 input, $50 output per million tokens. Double Opus 4.8. An enterprise running 10 agent workflows at 10,000 tasks per day lands north of $7,500 a day. Over $225,000 a month.

Uber’s CTO confirmed AI budget overruns on the previous model at half the price. One developer ran the math on moving his entire operation to Fable 5 at enterprise scale. The number came back in the hundreds of thousands a month. His reaction: you got any coupons?

The counterpoint: Fable 5 chews through complex jobs in fewer steps. A migration that takes a million tokens on a weaker model might take half that on Fable. Cost per task might be lower even though cost per token is higher. But that’s a calculation most teams haven’t had time to run.

The cost question filters harder than the capability question ever did. Capability is binary. Cost is a gradient. Most organizations hit the gradient first.

Wall Three: Availability

And then the government stepped in.

On June 12 at 5:21 PM ET, the US sent a letter ordering Anthropic to disable access for all foreign nationals. Everyone inside the US and outside it. Including Anthropic’s own overseas employees. In an official statement, Anthropic disputed the justification, noting the same capability exists in GPT-5.5 and open-weight models that remain unrestricted. They called it a misunderstanding about a national security directive.

Here’s what makes it strange. The government’s concern was that Fable 5 could be jailbroken to find security vulnerabilities. Anthropic reviewed the jailbreak demonstration and determined it found minor, already-known software bugs. The same technique works on OpenAI’s models — which were not recalled.

This is the first time a commercial AI product has been taken offline by government export controls. Previous restrictions targeted chips. This one targeted the model itself.

The third wall is the hardest. You can’t pay your way through a ban.

What Survives All Three

Five days before launch, on June 4, Anthropic published a warning about AI approaching recursive self-improvement — systems that get better at improving themselves without humans in the loop. They disclosed that Claude now writes over 80% of code merged into Anthropic’s own codebase, and engineer throughput is up 8x per quarter since pre-2025. They shipped Fable 5 anyway. Within 48 hours of going public, a jailbreak was demonstrated. Within 72 hours, the government shut it down.

The model is gone. The capability is real. The cost is known. And the week it all collapsed, not one of those walls was a technical limitation. They were economic and political.

What’s left is the human part.

Fable 5’s real breakthrough was autonomy. It could plan its own approach, hold entire projects in context, check its own progress, fix its own mistakes. That made it powerful. It also made it useless without someone pointing it in the right direction.

No model yet built can tell you what to work on.

That survives every wall. Direction. Judgment. Knowing that this problem matters more than that one. The week the model vanished, those skills didn’t. They can’t be regulated away or priced out.

You’ll be able to pay for a weaker model. You won’t have access to the strongest one. But knowing what to ask it to do is still on you. The models come and go. The questions you know to ask don’t.

The question isn’t which AI survives. It’s which human decisions do.

What are you working on right now that would survive a shutdown? And what happens when the next one comes along?

If this post resonated with you, buy me a coffee ☕ — it helps me continue sharing stories, ideas, and reflections.