The US Government Shut Down Claude Fable 5 in 3 Days
On June 9, 2026, Anthropic launched Claude Fable 5, its most capable public model. On the evening of June 12, a US export-control directive forced Anthropic to switch it off for every customer worldwide. It is the first time a government has used export controls to take down a live, broadly available AI API, and it happened within 72 hours of launch.
The precedent matters more than the model. Export controls, until now aimed at chips and model weights, have been applied to API access, on the strength of a disputed jailbreak report and without published technical findings.
What Fable 5 was
Fable 5 is the first public model in Anthropic's Mythos class, a tier above Opus. It shipped with a research sibling, Mythos 5, available as a restricted preview; Fable 5 is the version with extended safeguards for general use. Launch details:
- 80.3% on SWE-bench Pro, about 11 points ahead of the next-best model at launch
- 1M-token context window
- $10 per million input tokens and $50 per million output, twice Claude Opus 4.8
The safety design
Instead of refusing risky requests outright, Fable 5 runs automated checks on every request. When a request trips a classifier for offensive cybersecurity, biology, or extraction of the model's reasoning, it is silently handed to Opus 4.8. Anthropic's fallback guide says this happens in under 5% of sessions on average. For security teams the rate was far higher: one Claude Code issue reported that most of a defensive threat-intelligence session's messages came from Opus 4.8 after fallback.
The jailbreak claim
Shortly after launch, the red-teamer Pliny the Liberator announced that he had bypassed Fable 5's safeguards using Unicode homoglyphs, long-context framing, and task decomposition. Anthropic disputed that this was a real jailbreak, arguing that a meaningful bypass would have to expose broad cyber capability, not a narrow set of outputs.
The directive
At 5:21 p.m. ET on June 12, a letter from Commerce Secretary Howard Lutnick reached CEO Dario Amodei. Citing national security authorities and an "unacceptable risk" of diversion to military users in countries of concern, it required Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States, including foreign-national Anthropic employees.
Anthropic can't verify every user's nationality in real time, so it disabled both models for all customers. Other Claude models, including Opus 4.8 and Sonnet 4.6, stayed online.
Anthropic's response
Anthropic's public statement complied and pushed back in the same breath:
- It called the reported technique "a narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws."
- It said it had "not even received a disclosure of a concerning non-universal potential jailbreak that led to a harmful result."
- Its sharpest line: "If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers."
The irony is hard to miss. Anthropic has argued for years that governments need real authority over frontier model deployment. It got a version of that authority applied to itself, through export law rather than a transparent evaluation process.
What this establishes
- API access is now an export-control target, not just chips and weights.
- A single disputed jailbreak report can trigger a global shutdown within days of launch.
- Frontier launches carry political risk. A model can be pulled for reasons unrelated to reliability or commercial terms.
For builders, the lesson is operational. If your product depends on one frontier model, keep a tested fallback from another provider behind the same interface. Teams that could switch to Opus 4.8 or another vendor lost an afternoon.
Update: access began returning on June 26 for roughly 100 US companies and agencies. Commerce lifted the directive on June 30, and Anthropic restored Fable 5 on July 1.
Compare: Claude Fable 5 vs GPT-5.5 · Claude Fable 5 vs Claude Opus 4.8