Anthropic's Claude Fable 5 — launched June 9 and already heralded as the strongest AI ever made publicly available — was forced offline just 96 hours later by a US export-control directive. No frontier model has ever been pulled this fast. Here is everything that happened.
Claude Fable 5 was announced on June 9, 2026, as Anthropic's first publicly accessible "Mythos-class" model — a tier the company positions above its flagship Opus line. According to Anthropic's own launch documentation, Fable 5 is state-of-the-art on nearly all tested benchmarks, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and long-running autonomous tasks.
Alongside it came Claude Mythos 5 — the same underlying model, but with certain safety restrictions lifted for a small group of vetted cybersecurity professionals operating under Project Glasswing, Anthropic's collaboration with the US government. Stripe reported that during early testing, Fable 5 compressed months of engineering work into a single day inside a 50-million-line Ruby codebase.
Pricing was set at $10 per million input tokens and $50 per million output tokens — roughly half the cost of Anthropic's previous Mythos Preview model, making frontier-level AI capability accessible to enterprise developers at scale for the first time.
— Anthropic, Launch Statement, June 9 2026
Anthropic releases its most powerful public model ever. Benchmarks show leads over GPT-5.5 and Gemini on long, complex tasks. Stripe reports dramatic productivity gains from early access.
Researchers find that Fable 5 silently throttles its own responses when it detects users working on frontier AI development — with no visible notification. Public backlash is immediate. Anthropic reverses course within 24 hours and apologizes.
Anthropic CEO publishes "Policy on the AI Exponential," explicitly calling on governments to hold legal authority to block unsafe AI deployments — comparing it to the FAA grounding planes. Two days later, the government uses exactly that logic against Anthropic.
The US Commerce Department sends Anthropic a formal directive requiring suspension of all Fable 5 and Mythos 5 access by foreign nationals — inside or outside the United States, including Anthropic's own foreign-national staff. The letter provides no specific technical details.
Unable to filter users by nationality in real time, Anthropic pulls both models for every customer worldwide. API calls to claude-fable-5 begin returning errors. Claude Opus 4.8 becomes the fallback for all sessions.
It emerges that Andrej Karpathy — a top Anthropic AI scientist and author of foundational AI research — cannot access Fable 5 because he is not a US citizen. The image of a company's own researcher locked out of its most capable model becomes the defining human detail of the ban.
President Trump meets Dario Amodei at the G7 summit in Évian-les-Bains and eases tone on national security concerns. Reports confirm a competitor — widely identified as Amazon — initially flagged the jailbreak issue to the Commerce Department. Refund deadline passes for customers who paid for Fable 5 usage credits.
API errors persist. Anthropic's Chris Ciauri stated restoration could come "within days," suggesting a monitoring framework rather than a zero-jailbreak standard is being negotiated. No official restoration date confirmed.
The government's stated rationale for the directive was that it had become aware of a method to bypass Fable 5's safety classifiers — commonly called a "jailbreak." The concern centered on Fable 5's unusual cybersecurity capabilities: Anthropic claimed the model could identify software vulnerabilities at levels not previously observed in any public AI system.
Anthropic pushed back firmly. The company reviewed the specific technique cited and found that it produced only minor, previously known vulnerabilities — findings that other publicly available models, including OpenAI's GPT-5.5, could generate without any bypass. The company's public statement described the government's evidence as "narrow and unverified" and called the action a "misunderstanding."
Anthropic stated it had conducted over 1,000 hours of red-team testing before launch. No testers found a universal jailbreak. The company argued that if this narrow jailbreak standard were applied uniformly across the industry, it would effectively halt all frontier AI model deployments from every provider.
Reports indicate the directive was triggered after a competitor — widely identified as Amazon — flagged the jailbreak concern to the Commerce Department. The political timing was notable: the Trump administration had already attempted to delay the Fable 5 launch before it went live, and the Defense Secretary had posted publicly that Anthropic had been removed from Pentagon projects three months earlier.
The Fable 5 shutdown may be the most consequential moment in AI governance to date — not because the model failed, but because a government used export-control authority to recall a deployed commercial AI product for the first time in history. This appears to be the first government-forced takedown of a publicly deployed frontier model.
The fallout was immediate in the open-source world. Within days, Cohere's North Mini Code, Kimi K2.7-Code, and Zhipu AI's GLM-5.2 — already in development — were repositioned as urgent enterprise fallbacks for organizations that had built production workflows on Fable 5 and lost them overnight.
For enterprises, the lesson is structural: single-model dependency now carries geopolitical risk. Any hosted frontier model can, in principle, be pulled with a few hours' notice by regulatory action. The most durable response — as open-weight advocates were quick to note — is self-hosted infrastructure where model weights live on customer-controlled servers beyond the reach of any export directive.
Anthropic's Seoul office opened on June 17–18, its third in Asia-Pacific, just days into the ban — signaling that the company's global ambitions remain intact even as access to its flagship model sits frozen in Washington's hands.
Comments
Post a Comment