US Government Forces Anthropic to Pull Fable 5 and Mythos 5

At 5:21pm ET on June 12, a letter from the US Commerce Department landed on Anthropic's desk and, within hours, two of the most capable AI models on the planet went dark for everyone. In a public statement, Anthropic said it was disabling Claude Fable 5 and Claude Mythos 5 for all customers to comply with an export-control directive issued under national-security authorities — a directive the company says arrived with no specifics attached. "Access to all other Anthropic models will not be affected," the statement noted, but the flagship tier is, for now, simply gone.

This is the first time a US administration has reached in and pulled a commercially deployed frontier model off the market. It happened three days after launch.

Who ordered it

The order did not come from an anonymous corner of the bureaucracy. According to Axios, which broke the story, Commerce Secretary Howard Lutnick sent the letter directly to Anthropic CEO Dario Amodei, placing Fable 5 and Mythos 5 under export controls covering any location outside the US and "all foreign persons within the country" — a scope sweeping enough to include Anthropic's own foreign-national employees. That breadth is what forced a full shutdown rather than a geofence: with no clean way to wall the models off from foreign persons already inside the US, disabling them for everyone was the only compliant move.

An administration official told Axios the Commerce Department acted after another company claimed it had jailbroken Mythos, which alarmed officials about national-security risk. The official said the administration first tried to get Anthropic to pause the launch, and when that failed, the export-control letter followed. Anthropic has not named the company that made the jailbreak claim, and neither has the government.

The jailbreak at the center of it

Here is the detail developers will find almost comic: the "method of bypassing" Fable 5 that triggered a federal intervention was, per Anthropic, asking the model to read a specific codebase and fix any software flaws it found. That's it — the core loop of half the tools in modern software development.

Anthropic says it reviewed a demonstration of the technique and watched it surface "a small number of previously known, minor vulnerabilities," all "relatively simple" and discoverable by other publicly available models — it specifically pointed at OpenAI's GPT-5.5. The company stressed that Fable went through thousands of hours of red-teaming with the US government, the UK's AI Safety Institute, and third parties, and that no universal jailbreak was found.

Its objection is blunt. Anthropic said it "disagree[s] that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people," and warned that if this standard were applied across the board, "it would essentially halt all new model deployments for all frontier model providers." In other words: by this bar, nobody ships another frontier model.

The irony isn't lost on anyone. Fable 5 was already the safety-restricted configuration — the same base weights as Mythos 5 but wrapped in heavier guardrails, the version Anthropic positioned as the responsible one. It got pulled anyway.

The model that was already under fire

Fable 5 spent its three days of life as one of the most divisive releases in recent memory. At launch, Andrej Karpathy called it "a super exciting release… SOTA on everything by a margin" and "a major-version-bump-deserving step change forward." The benchmarks were not in dispute.

The guardrails were. Researchers spent the week arguing the safety layer had been tuned so aggressively it broke legitimate work. "The claude fable 5 nerf for AI research has induced the angriest reaction from AI researchers that I've ever seen in my life," said researcher Ethan Caballero. Developers reported routine security and code-review tasks getting flagged as if they were attacks; one founder called the restrictions "deeply unethical and disrespectful to developers and scientists."

And the jailbreaks the government is worried about? Some were already public. Prolific red-teamer Pliny the Liberator had announced bypassing Fable 5's safety layers within days of release using a multi-agent attack, and reportedly leaked the model's roughly 120,000-character system prompt to GitHub. The cat the directive aims to contain was, by most accounts, already out of the bag.

The developer reaction

Once the suspension hit, the anger curdled into something closer to a wake. Theo (t3.gg), who'd posted earlier in the week that he'd burned through more than $1,000 of tokens in a single day on his $200 plan, wrote a mock-elegy: "Fable, my beloved, I will miss you so. Our three days together were magical." It was a joke. It was also, for a lot of teams who'd wired the model into real workflows over the weekend, not really a joke at all.

The structural worry underneath the memes is the one Anthropic flagged itself: if a single demonstrated jailbreak — especially one as mundane as "find bugs in this repo" — is grounds for a federal recall, the precedent reaches every lab. Today it's Fable and Mythos. The standard, once set, doesn't care whose logo is on the model.

What happens now

Anthropic says it is complying while it disputes the order, and that all other Claude models remain available. The company hasn't given a timeline for whether Fable 5 and Mythos 5 come back, under what conditions, or whether a narrower geofenced version might satisfy Commerce. The government hasn't elaborated on the specific national-security concern it declined to detail in the letter.

For developers, the immediate reality is simpler and stranger: the best models Anthropic has ever shipped existed publicly for 72 hours, and then the US government decided you couldn't use them. We'll update this story as Anthropic, Commerce, and the labs watching nervously from the sidelines respond.

Sources: Anthropic's official statement; Axios (scoop, Commerce Secretary Lutnick letter); CNBC; Bloomberg; 9to5Mac.