US Government Pulls Anthropic’s Two Most Powerful AI Models – 3 Days After Launch

📖 5 min read

At 5:21 PM ET on Friday, June 12, 2026, Anthropic received a letter from the US government. By midnight, its two most powerful AI models were dark for every customer worldwide.

The US government issued an export control directive ordering Anthropic to immediately suspend all access to Claude Fable 5 and Claude Mythos 5, models the company had released just 3 days earlier. Citing national security concerns, the directive technically applied to “any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.” Anthropic’s response: rather than build nationality-verification infrastructure overnight, it shut the models down for everyone.

This is the first time in history a frontier AI model has been pulled from market by a government order after public release.

What the Government Said – and What Anthropic Heard

The letter provided no specific evidence. Anthropic’s statement says the government verbally indicated it had discovered a method to “jailbreak” Fable 5 – and that this jailbreak essentially amounted to asking the model to read a codebase and identify software vulnerabilities.

Anthropic was blunt in its disagreement: “We reviewed a demonstration of this specific technique being used to identify a small number of previously known, minor vulnerabilities. These vulnerabilities all appear relatively simple, and we have found that other publicly-available models are able to discover them as well without requiring a bypass.”

Translation: GPT-5.5 and other widely-available models can do the same thing, without any jailbreak.

The company had spent thousands of hours red-teaming Fable 5 in collaboration with the US government, the UK AI Safety Institute, and private third-party organizations before launch. Their safeguard report found Fable’s defenses “substantially more effective than those of any previously deployed model” and noted no universal jailbreak – a method that broadly unlocks dangerous capabilities – had been discovered.

Anthropic’s position: “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.”

This Is Not Anthropic’s First Fight With Washington

The shutdown doesn’t exist in a vacuum. Earlier in 2026, Trump’s Department of Defense labeled Anthropic a “supply chain risk” – a designation that effectively barred government agencies and contractors from using Claude. Anthropic responded by filing lawsuits against the Trump administration, a remarkably aggressive move for a company that has historically positioned itself as the responsible, safety-first lab.

The irony is hard to miss: Anthropic’s unusually transparent safety warnings about Fable’s power – the company had called Mythos “too powerful to release” before adding safeguards for its Fable variant – may have made the model a more visible target for government scrutiny. A quieter rollout, like OpenAI’s approach with GPT-5.5, faced no similar order.

The Fallout: Who Got Hit and How Bad

The immediate casualties were significant:

• Cognition/Devin – the AI coding agent – removed Fable 5 from its platform the same night
• Agent Arena pulled both models from its benchmarks
• Anthropic reset its 5-hour and weekly rate limits as a partial gesture toward affected users
• Artificial Analysis noted it was “the first time our Intelligence Frontier chart has moved backward”

The scope: Fable 5 and Mythos 5 had launched to Anthropic’s entire user base, which the company has described as “hundreds of millions of people.” Developers who had built products on the API, customers paying for Claude Pro, and enterprise users all lost access to the two top-tier models with no warning and no recovery timeline.

What This Means for Everyone Using AI

The Fable/Mythos shutdown crystallized a risk that infrastructure and product teams have been quietly discussing: frontier AI API access can vanish overnight, and for reasons that have nothing to do with technical failures.

Engineers and AI researchers quickly reframed the story as a “model sovereignty” problem. The practical concern: if you build on a single closed frontier API – whether from Anthropic, OpenAI, or Google – you are exposed to geopolitical risk, export controls, and government mandates you have zero control over. The models don’t have to fail. The government just has to call.

This is precisely the argument open source AI advocates have been making. Within hours of the announcement, open-source AI supporters were trending under the banner that owning your stack is the only real hedge. Cohere, which operates its own models and infrastructure, pointed out the obvious: “owning the stack matters.”

Benchmark and Business Impact

The performance stakes are real. Before the shutdown, Claude Fable 5 paired with Claude Code had just entered Artificial Analysis’s Coding Agent Index at rank 1 with a score of 77, narrowly ahead of Codex + GPT-5.5 at 76. That ranking is now moot – the model is unavailable, and the leaderboard has a gap at the top.

Agent + Model	DeepSWE Score	Status
Claude Code + Fable 5 [max]	77	Suspended
Codex + GPT-5.5 [xhigh]	76	Available
Claude Code + Opus 4.8 [max]	73	Available

For Anthropic, the business cost is compounding. The supply chain risk designation already blocked government contracts. The lawsuits are ongoing. Now a public product with users baked in has been forcibly recalled. The company’s revenue trajectory for 2026 – at a $61.5 billion valuation, it needs growth – just took a visible hit.

The Open Questions

As of June 16, several critical things remain unknown:

• Timeline for restoration: Anthropic has not provided one. The directive has no stated expiration.
• Whether the jailbreak is real: Anthropic disputes the severity. The government has not published evidence.
• Legal next steps: Anthropic is already in litigation with the Trump administration. This directive is almost certainly being reviewed by Anthropic’s legal team.
• What happens to Mythos customers: The enterprise-tier Mythos class had a separate data retention policy (30 days) that gave Anthropic jailbreak monitoring capability. Those customers are also locked out.

BetOnAI Verdict

The technical case against the shutdown is strong. A “narrow, non-universal jailbreak” that enables capabilities also present in GPT-5.5 is not a Fable-specific catastrophe – it’s a baseline problem for the entire industry. If the government’s threshold for pulling a model is “a jailbreak exists,” every frontier model currently deployed should be offline.

But the political context matters. Anthropic made enemies in Washington by refusing to let the military use its technology without restrictions, then sued the DoD when they fought back. When you’re already labeled a supply chain risk and your lawsuit is active, the government isn’t going to give you the benefit of the doubt on a borderline safety call.

The real lesson here is infrastructure. If you’re building products or workflows that depend on any single closed-API provider – Anthropic, OpenAI, or Google – the Fable/Mythos incident is your stress test. You just saw the failure mode. The solution is either multi-provider redundancy or open-weight models you control. The latter option got a lot more attractive this week.

For Anthropic: they’re right on the technical merits and losing on the political ones. Whether that changes depends on the courts, the next administration, or whether this particular fight costs them enough customers to negotiate a different outcome.

---

Sources:

• Anthropic Official Statement – Jun 12, 2026
• WIRED – Anthropic Says It’s Taking Claude Fable 5 Offline
• TechCrunch – Anthropic’s safety warnings may have backfired
• BBC News – Anthropic’s Claude Fable 5 and Mythos 5 AI suspended
• Latent Space AINews – Fable and Mythos officially too dangerous to release