The models are back. On Tuesday evening Anthropic said the US Department of Commerce had lifted the export controls that forced it to switch off Claude Fable 5 and Mythos 5 less than three weeks ago, and that access would begin returning the next day. Commerce Secretary Howard Lutnick confirmed the reversal, saying the two sides had "worked closely" to "strengthen America's leadership in AI." It closes one of the stranger episodes in the short history of frontier AI regulation: a government reaching into a private company's product line, pulling its two most capable models off the market worldwide, and then, after a fortnight of negotiation, letting them back on.
The terms of the truce say a lot about where the line now sits. Anthropic agreed to proactively detect and address security risks, to work with the government on release protocols for Mythos, Fable and future models, and to flag any malicious activity. Lutnick reserved the right to reimpose controls "should circumstances change." The original trigger, by several accounts, was a jailbreak of Fable 5 reported by a trusted partner later identified as Amazon: a way around the guardrails on a model built for public use. Anthropic maintained the flaw was narrow, surfaced only previously known vulnerabilities, and could be reproduced on other public models including OpenAI's. It has since shipped a safeguard that blocks the specific technique.
If the reversal was the headline, the more revealing move came the same week with a different model. Anthropic released Claude Sonnet 5, a cheaper workhorse it says matches the far pricier Opus 4.8 across many benchmarks at less than half the cost. What the company chose to advertise, unusually, was a weakness: Sonnet 5 performs "substantially poorer" on cybersecurity tasks than Opus 4.8 or Mythos 5. "We did not deliberately train Sonnet 5 on cybersecurity tasks," Anthropic wrote, noting that a partial success against Mozilla's Firefox was "likely due to improvements in general intelligence rather than specific training." A lab boasting about what its model cannot do is a sign of the new weather.
That is the quiet story underneath the noise. The mechanism that briefly grounded Fable 5 is not going away. Under an executive order signed this month, developers are asked to hand "covered frontier models" to the government for up to 30 days before release, and the same week Anthropic was cleared, OpenAI delayed the full launch of GPT-5.6 at Washington's request. Isaac Harris of the Frontier Security Institute read the outcome as proof that "there's now a process for standards of the US models." Anthropic, which the Pentagon separately labelled a "supply-chain risk" this year after it refused to let its models be used for mass surveillance or autonomous weapons, appears to be building to that process: powerful where it is safe to be, deliberately dull where capability invites a fight.
The open question is what the process actually buys. Harris put his finger on it, warning there is "still a question mark as to how equivalently dangerous capabilities coming from China with less guardrails will be handled." During the three weeks Fable 5 sat dark, the most capable model many users outside America could legally reach was a free download from Beijing's Z.ai, a company on Washington's own trade blacklist. The controls can shape what a US lab ships, and when. They have far less to say about what everyone else builds, or what quietly leaks out around the edges. That gap is the subject of a second story in today's issue.