The US Government Shut Down Anthropic's Most Powerful Model in 3 Days

On June 9, 2026, Anthropic launched Claude Fable 5 — the company's most capable model ever released to the public. By the evening of June 12, it was gone. A US Department of Commerce export control directive forced Anthropic to disable the model worldwide, marking the first time a government has used export controls to pull the plug on a live AI API.

Here's what happened, why it matters, and what it means for the future of frontier AI deployment.

What Was Fable 5?

Claude Fable 5 was a "Mythos-class" model — a new tier above the Opus family, representing Anthropic's fifth generation of Claude. It shipped alongside Claude Mythos 5, which shared the same underlying weights but with most safety guardrails removed. Mythos 5 was restricted to roughly 150 vetted organizations under Anthropic's "Project Glasswing" program.

The numbers were staggering:

• SWE-bench Pro: 80.3% (vs GPT-5.5's 58.6%)
• FrontierCode Diamond (agentic coding): 29.3% (vs Opus 4.8's 13.4%)
• ExploitBench (cybersecurity): 78.0% (vs Opus 4.8's 40.0%)
• GDPval-AA (knowledge work): 1932 (vs GPT-5.5's 1769)

Pricing was aggressive: $10 per million input tokens, $50 per million output tokens — roughly 2× the cost of Opus 4.8. Stripe reportedly used it to migrate a 50-million-line Ruby codebase in a single day.

The Safety Architecture

Rather than hard-blocking dangerous queries, Fable 5 introduced a classifier-based fallback system. When a request tripped a safety classifier — for domains like cybersecurity, biology, chemistry, or model distillation — the query was silently rerouted to the weaker Claude Opus 4.8. Anthropic said this fallback triggered in less than 5% of sessions.

This design was meant to let Fable 5 remain broadly useful while containing the highest-risk capabilities. It was also, as it turned out, the mechanism that would be exploited.

The Jailbreak

Within 48 hours of launch, a red-teamer known as "Pliny the Liberator" — head of collective BT6 (Basilisk Team Six), named to TIME100 AI 2025 — claimed to have bypassed Fable 5's safety classifiers. The method used what Pliny called a "pack hunt": a multi-agent attack combining Unicode substitutions, narrative framing, and task decomposition to slip past the fallback routing.

Screenshots circulated showing Fable 5 producing stack buffer overflow exploit code and chemical synthesis pathways. Pliny also leaked the model's 120,000-character system prompt to GitHub. Fortune, The Register, and NBC News all ran stories within hours.

The Government Steps In

On June 12 at 5:21 PM ET, Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei. The directive invoked national security authorities under the Export Administration Regulations (EAR), ordering Anthropic to block "all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees."

Because Anthropic cannot reliably verify user nationalties in real time across its API, the practical result was a total global shutdown of both models. Every customer, in every country, lost access simultaneously. Other Anthropic models — Opus 4.8, Sonnet 4.6, Haiku — remained online.

Anthropic's Response

On June 13, Anthropic published a public statement pushing back hard. Key points:

• The jailbreak was "narrow and non-universal" — it found a small number of previously known, minor vulnerabilities
• Other publicly available models (including GPT-5.5) can discover the same vulnerabilities without any bypass
• The demonstrated capability is "widely used daily by cybersecurity defenders"
• The government letter provided no specific technical details — only verbal evidence
• The technique "essentially consists of asking the model to read a specific codebase and fix any software flaws"

Anthropic's most pointed statement:

"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

What makes this episode remarkable is that Anthropic had actively advocated for government oversight of frontier AI. Just days before the launch — on June 5 — Dario Amodei published articles calling for the government to have authority to block model deployments based on third-party evaluations. Anthropic had pushed for regulatory thresholds tied to training compute.

The government used that very authority, but in a far broader and more aggressive form than Anthropic envisioned. Instead of a transparent, evaluation-based process, it was a verbal directive with no written technical justification — and the net effect was a global kill switch.

What This Means

This is a watershed moment for AI regulation. Several things are now established:

• Export controls can target API access, not just chip sales or model weight downloads. This is a new category of enforcement.
• A single jailbreak report — even one the developer disputes — can trigger a complete model recall within 72 hours of launch.
• Frontier model launches carry political risk that didn't exist six months ago. Every new capability tier is now a potential target.
• The "responsible AI" position doesn't protect you from your own regulatory asks being turned against you.

Both Fable 5 and Mythos 5 remain disabled as of this writing. Anthropic says it is "working to restore access as soon as possible." There is no timeline.

\n

\n At lmrank.com, we track live scores and benchmark results for frontier AI models. For the latest on Claude Fable 5's status and alternatives, see the model page.