SAN FRANCISCO — Three weeks ago, the United States government decided that Anthropic’s most capable AI model was too dangerous for the rest of the world to touch. On Tuesday, it decided the opposite, and the explanation for the reversal is more damaging to the original decision than the ban itself ever was to Anthropic.
The Commerce Department’s Bureau of Industry and Security notified Anthropic that export controls on Claude Fable 5 and Mythos 5 had been lifted, and the company began restoring global access on Wednesday, July 1, across the Claude Platform, Claude.ai, Claude Code and Claude Cowork. NBC News reported Commerce Secretary Howard Lutnick posted on X that his department had “worked closely with Anthropic to analyze and approve Fable 5,” and followed with a letter to Anthropic co-founder Tom Brown confirming the decision. The tone was cooperative. The underlying story is that the government’s national-security case for banning the model in the first place did not survive its own testing.
The models were pulled offline on June 12 after Amazon researchers found a way to prompt Fable 5 into producing exploit code for a known software vulnerability, a finding Lutnick’s Commerce Department treated as evidence the model could hand cybercriminals a meaningful new capability. Anthropic disputed that reading immediately, and more than 80 cybersecurity executives, some of whom had spent the preceding weeks warning enterprises to brace for Mythos-class threats, signed a public letter arguing the capability Amazon had found was not unique to Fable 5 at all.
Anthropic’s own account of the reversal, posted Tuesday, confirms the security executives were right. According to the company, testing showed that Claude Opus 4.8, OpenAI’s GPT-5.5 and Moonshot’s Kimi K2.7 could all reproduce the identical exploitation demonstration Amazon had flagged in Fable 5. Every model tested produced the same output. The technique, Anthropic said, represented “routine defensive cybersecurity work” rather than a capability unique to its most powerful system. The Department of Commerce’s own Center for AI Standards and Innovation tested Anthropic’s new safety classifier, which blocks the specific bypass in more than 99 percent of cases, and confirmed the safeguards were “extraordinarily strong.” What went unexplained, in both Lutnick’s statement and Anthropic’s, is why a capability present in three competing models justified restricting only one of them for three weeks.
The price of getting the ban lifted looks steeper than the ban itself. Anthropic committed to four pillars of expanded cooperation with the government: pre-release access and evaluation for models that advance national security capabilities, rapid information-sharing on jailbreaks and safeguard responses, dedicated joint research teams, and common industry standards developed with government input. None of that existed in this form before June 12. A company that spent its first years marketing independence from military and surveillance use cases has now agreed to hand federal reviewers a look at unreleased models before the public sees them, in exchange for the right to sell the ones it already built.
Mythos 5, the more capable and less publicly available of the two models, was not fully freed. Its access remains limited to roughly 100 vetted U.S. organizations inside Anthropic’s Glasswing partner program, and the company says it is still coordinating with the government on wider release. Fable 5 fares better: full global availability across Claude’s consumer and developer products, with Pro, Max, Team and select Enterprise subscribers getting up to half their weekly usage on the model at no extra cost through July 7.
Not everyone in the industry treated the outcome as a clean resolution. OpenAI’s Sam Altman described the government’s practice of granting phased, case-by-case approval for advanced model access as “bad news,” a rare public jab from a rival that generally avoids commenting on Anthropic’s regulatory entanglements. Altman’s complaint is less about Anthropic specifically than about the precedent: if Washington can freeze and unfreeze a frontier model on a three-week timeline based on evidence it later has to walk back, every lab now has to plan for that as a standing operational risk, not an exceptional one.
Anthropic is also still in federal court over the underlying directive, after at least one customer sued the government in June arguing the export control caused direct commercial harm, and that litigation does not automatically end because the restriction it challenges has been lifted. Whether Anthropic presses the case to establish a legal check on future emergency export controls, or lets it go now that the practical injury has stopped, will say more about how much the company trusts this arrangement to hold than anything in Tuesday’s announcement.
What the resolution does not answer is what happens the next time a red-teamer, at Amazon or anywhere else, finds a bypass in a frontier model and hands it to a government that has just demonstrated it will act on incomplete comparative testing. The safeguard is real. The four pillars of cooperation are real. But the sequence, ban first, compare capabilities across competitors later, is now a precedent of its own, and nothing announced this week prevents it from happening again to Anthropic or to whichever lab ships the next model that scares a regulator before anyone checks whether the fear was specific to that model at all.
