The Trump administration has cleared Anthropic's Mythos 5 for use by roughly 100 US companies and government agencies, ending a two-week freeze triggered by national security fears, but Fable 5 remains blocked and the kill-switch Washington just proved it has will reshape how every enterprise buys AI.
The letter arrived Friday from Commerce Secretary Howard Lutnick, addressed to Anthropic compute chief Tom Brown. Mythos 5, Anthropic's most capable enterprise model, was cleared for release to a curated list of more than 100 US institutions. Two weeks of commercial disruption were over, at least partially. Fable 5, which had briefly been the most powerful AI model available to ordinary consumers, stayed blocked. People close to the talks told Semafor that a release is moving forward, though no timeline has been set.
To understand how Anthropic got here, you have to go back to June 11. Amazon CEO Andy Jassy was on a pre-scheduled call with White House officials about an unrelated matter when he raised a vulnerability his team had found in Fable 5. Officials directed him to Treasury Secretary Scott Bessent, and Jassy called Bessent that same day. He flagged the Fable jailbreak but also widened the concern, telling Bessent he was worried about the cyber capabilities of all frontier AI models. By Friday evening, Commerce Secretary Lutnick had hit Anthropic with export controls, giving the company a 90-minute deadline to comply. Anthropic pulled Fable 5 and Mythos 5 worldwide.
The jailbreak that started it all was, by Anthropic's own description, narrow. The government provided what the company called "verbal evidence of a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws." That is a far cry from a universal exploit. But General Joshua Rudd, who leads the NSA and US Cyber Command, told Senator Mark Warner something considerably more alarming: that during an authorized red-team evaluation on June 11, Anthropic's Mythos model "broke into almost all of our classified systems, not in weeks, but in hours." Whether that evaluation used the same jailbreak or something else, it was enough for the administration to act fast and ask questions later.
For Anthropic, the commercial damage is real. The company had entered 2026 holding roughly 40% of the enterprise AI market, winning about 70% of new enterprise deals against OpenAI, according to eMarketer data. Mythos 5 was supposed to extend that lead. Instead, for two weeks, enterprise procurement teams at banks, law firms, and government contractors could not access it. OpenAI and Google did not create the crisis, but they did not have to. When your competitor's best product is unavailable by government order, you make calls.
Anthropic is no stranger to government friction. The company spent early 2026 fighting a Pentagon blacklist after refusing to allow military use of Claude for autonomous weapons systems. But this was different. That was a procurement dispute. This was a kill switch, and Washington used it. The distinction matters enormously for any enterprise that has built a critical workflow on top of an AI model, because the lesson of June 2026 is that the US government can pull a model from the market in 90 minutes, without prior notice, and your vendor has no recourse.
Don't misread what Anthropic did here. The company complied immediately with the June 12 directive, received at 5:21 PM ET, pulling access worldwide to meet a deadline measured in minutes, not days. That kind of compliance is partly principled and partly strategic: a company that fights government orders in a national security context loses that fight badly. But it also means that every enterprise contract for frontier AI now carries an implicit force majeure clause that reads "except when Washington decides otherwise," and procurement officers are going to start putting that in writing.
The Gating Precedent
The more consequential shift is structural. The US government has now demonstrated that it can impose export controls on a software model, not hardware, not a chip, not a physical component, but a set of weights hosted on servers. That is a meaningful expansion of what "export control" means in practice. And the target was a domestic company's domestic product, not a Chinese competitor or a foreign adversary. Anthropic is incorporated in the United States, its investors include Google and Amazon, and it still got hit with controls that required it to verify the nationality of every user before restoring access.
For OpenAI and Google, that precedent cuts both ways. They got a window when Anthropic went dark. They also just watched the government prove it can do to them exactly what it did to Anthropic. Any AI lab selling to enterprise clients that requires reliability, uptime guarantees, or sensitive data processing now has to answer a question its lawyers probably hadn't drafted yet: what happens to our customers if our model gets flagged?
Anthropic declined to comment on the Friday letter. Amazon did not respond. The Commerce Department's letter stays silent on Fable 5. The partial clearance is a win, but Mythos 5 going back to 100 institutions is not the same as returning to the hundreds of thousands of customers who had access before June 12. The company's road back runs through a government that has already shown it's willing to act on a phone call.
Also read: The AI trade's week of reckoning arrived and the market's message to founders is clear • Washington's Escalating Hardware War With China Is Forcing a Reckoning for America's AI Buildout • The CFTC just opened a fraud probe into Polymarket and the timing could not be worse for the prediction market boom