Why Claude Fable 5 vanished days after launch

Fable 5 was new. Anthropic released it on Tuesday, June 9. It was the first time the company put a model this strong in front of the public. Until then, its most capable work sat behind closed doors.

Fable came from a bigger family. Anthropic had built a research model called Mythos Preview. It was good at reading code and finding security flaws. A small group of companies tested it through a program called Project Glasswing. Mozilla said the model helped it fix hundreds of vulnerabilities.

From that base, Anthropic made two models. Mythos 5 stayed locked to the Glasswing partners. Fable 5 went public, but with heavy guardrails. The safeguards blocked the model in high risk areas like cybersecurity and biology. Some users even said the limits were too strict.

It was not a quiet release. By benchmark tests from Vals AI, Fable 5 was the most capable model anyone could use. The price matched its level. Ten dollars per million input tokens. Fifty dollars per million output tokens. Builders were told to try it early. Four days later, that advice aged badly.

The order came on Friday, June 12, at 5.21 pm Eastern time. It was an export control directive from the US government. Commerce Secretary Howard Lutnick sent the letter to Anthropic's CEO, Dario Amodei. Narrow on paper. Wide in effect.

The government told Anthropic to block any foreign national from Fable 5 and Mythos 5. That covered people outside the United States and foreign nationals inside it. It even reached Anthropic's own foreign born staff. The letter cited national security. It did not spell out the technical concern.

So Anthropic made a hard call. It could not cleanly fence off foreign users from everyone else. To comply, it shut both models down for all customers. Every other Claude model stayed online. By Friday night, the Fable 5 page simply said the model was temporarily unavailable.

Anthropic's read on the order points to one thing. A jailbreak. The company believes the government saw a way around Fable's safeguards. The method, as Anthropic describes it, is plain. You ask the model to read a codebase and fix the flaws it finds.

Anthropic looked at a demonstration tied to the order. It found the model had surfaced a few already known, minor bugs. Nothing exotic. The company says other public models do the same with no bypass at all. It pointed to OpenAI's GPT-5.5 as one example. It also noted that security defenders use this kind of work every day.

Reporting from Axios said the action followed a claim by another company that it had jailbroken Mythos, which worried officials. The same report said the administration had tried to stop the release earlier and could not. The government acted on a national security concern. Anthropic says the evidence it has seen so far was verbal, not a documented universal jailbreak.

One model is not the real lesson here. The risk is building on a single provider with no backup. Friday showed how fast access can vanish. No outage. No bug. A letter, and the most capable public model was gone within hours.

If you build on closed AI, treat provider risk as real. Here is what that looks like in practice.

There is a sharper edge for readers outside the United States. The order blocked foreign nationals first. If you build from Dhaka, Delhi, Dubai, or anywhere outside US borders, you sit on the wrong side of that line by default. The best models may start reaching you later, or not at all. That is a planning problem now, not a someday problem.

For years, the fight over AI power ran through hardware. Chips. GPUs. Cloud capacity. The US controlled who could buy the fastest silicon. Models felt like the open part of the stack.

Friday changed the shape of that. This looks like the first time the US government took a publicly deployed model offline. Not a chip. Not a data center. The model itself. Once a government can do that once, it can do it again.

That raises a question every builder should sit with. If the best models get tied to where you live and the passport you hold, access stops being something you buy. It becomes something you are granted. That is the real shift to watch.

Anthropic expects this to be temporary. Maybe it is. The directive could lift in days. But the precedent will outlast the outage. The door is open now. The question is who walks through it next.