Two weeks ago the US government reached into a private company's product line and switched it off. On Friday it switched part of it back on, but with conditions that may outlast the crisis that produced them. In a letter to Anthropic, Commerce Secretary Howard Lutnick lifted the export block on Claude Mythos 5, the company's most powerful model, and cleared it for release to more than 100 American institutions, including major companies and government agencies. The same afternoon, OpenAI shipped GPT-5.6 to its own short list of government-approved partners. The standoff is easing. What it leaves behind is something new: a working system for licensing who may use a frontier AI model, assembled in a fortnight and unlikely to be dismantled.
The mechanism is worth reading closely, because it borrows its language from arms control. "A license will no longer be required to export, reexport, or in-country transfer the Claude Mythos 5 Model to entities identified in Annex A to this letter," Lutnick wrote, naming an approved list and extending it to those companies' foreign-national employees and to Anthropic's own. In plain terms, the model weights are now treated like a controlled good: free to move among named, trusted parties and restricted everywhere else. Two weeks ago Commerce had invoked export law to pull Mythos and its lighter cousin Fable 5 entirely offline, arguing that even Anthropic's foreign staff needed a federal license to touch them. Friday's letter does not repeal that logic. It operationalises it.
The trigger for all of this was capability, not paperwork. Commerce moved against Mythos after warnings from Amazon and others that the model could be "jailbroken" for malicious use, and the specific fear was cyber: a system able to find software flaws and run multi-step intrusions with little human oversight. The administration says it acted fast to contain that. "In just two weeks, we have worked diligently to ensure America remains the global leader in AI while safeguarding our security," a Commerce spokesman said. Lutnick's letter credits "significant progress" in daily talks, and notes that Anthropic "has committed to work with the U.S. government on protocols and standards and releases" for its models. That last phrase is the quiet part: the company has accepted government review as a permanent feature, not a one-off.
The people left outside Annex A are noticing. The letter says nothing about Fable 5, the weaker model that was briefly the most powerful AI widely available to consumers, and nothing about when, or whether, anyone abroad gets access. European officials and other US allies have already complained about a new dependence on decisions made in Washington. In India the exclusion has landed as a strategic jolt: commentators there note that Indian IT firms are now cut off from Fable 5's coding ability even though Indian data helped train it, and have begun calling the episode a "kill switch" moment that argues for building sovereign models rather than renting American ones. A control regime meant to reassure allies is, for the moment, mostly reminding them how little say they have.
The deeper problem is that the regime is being built faster than anyone can say what it is for. There is still no published standard for what makes a model too dangerous to release, no settled agency in charge, and no statutory backing beyond export law stretched to cover software. Friday's de-escalation proves the system can move quickly when a powerful company and a nervous government both want a deal. It does not prove the same machinery would be fair, predictable, or even lawful applied to a smaller lab with less leverage. Washington has shown it can license a frontier model in two weeks. The harder question, who gets told no and why, is still unanswered.