AI companies launch a new "best ever" model most weeks now, and most of those launches blur together. This one did not.
Anthropic shipped a brand new top-tier Claude model in June 2026. Three days later, the US government told Anthropic to stop giving people access to it. Not a bug fix, not a quiet rollback, an actual government order pulling a commercial AI model that was already in use by hundreds of millions of people.
What actually happened
On 9 June 2026, Anthropic launched Claude Fable 5, a new flagship model, alongside a more locked-down sibling called Mythos 5. Anthropic described Fable 5 as "a Mythos-class model that we've made safe for general use," which is Anthropic's way of saying Fable 5 has roughly the same underlying capability as Mythos 5, but with extra guardrails so it can be handed to the public.
Three days after launch, on 12 June 2026, Anthropic published a statement saying the US government had directed it to suspend access to both models. The reason: a possible way to get around the safeguards meant to stop the model being used for serious cybersecurity work, things like finding and exploiting software vulnerabilities at a level beyond what an everyday user should be able to do with a chatbot.
Anthropic's own account of how this "jailbreak" (industry shorthand for tricking an AI model into ignoring its safety rules) actually worked is fairly mundane on the surface: someone could ask the model to "read a specific codebase and fix any software flaws." The government's concern was that a request phrased that innocently could still be used to get around the safeguards meant to limit access to Mythos-level cybersecurity capability.
Anthropic pushed back. It said in its own statement that a narrow, non-universal jailbreak "should not... be cause for recalling a commercial model deployed to hundreds of millions of people." It suspended access anyway, in line with the order.
That suspension lasted roughly three weeks. Anthropic's own announcement page confirms access to Fable 5 and Mythos 5 was restored on 1 July 2026.
Anthropic did not slow down after that. It launched Claude Sonnet 5 on 30 June 2026, calling it "the most agentic Sonnet model yet" and making it the default model for Free and Pro plan users, priced at $2 per million input tokens and $10 per million output tokens through 31 August 2026 (rising to $3/$15 after that date). Then on 24 July 2026, Anthropic launched Claude Opus 5, pitched as near-frontier performance at half the cost of Fable 5, priced at $5 per million input tokens and $25 per million output tokens. Anthropic's own cybersecurity testing found Opus 5 is about as good as Mythos 5 at spotting vulnerabilities in code, but "considerably less successful" at actually turning those vulnerabilities into working exploits, a gap Anthropic says is deliberate.
Why it actually matters
For most people using Claude day to day, this changes very little. Anthropic said the suspension had no reported impact on its other, non-frontier models, so anyone using Claude for writing, research, or everyday questions would not have noticed anything.
What it does signal is a shift in who gets the final say on whether a new AI model is safe enough to release. This is the second time in a matter of weeks that a frontier AI lab has had a model release delayed or pulled by a government over cybersecurity concerns, following a similar story involving OpenAI's GPT-5.6. Increasingly, that call is not just an AI company's to make.
What's still unclear
Anthropic's own statement says the government directive targeted access "by any foreign national, whether inside or outside the United States." In practice, though, the suspension affected access globally, not just for foreign nationals, so the real-world effect was broader than the directive's stated wording suggests. It is also worth noting that Anthropic's Opus 5 announcement does not give the model a formal AI Safety Level classification, so that detail is left out here rather than filled in from outside reporting that could not be independently confirmed.
