The Commerce Department has let Anthropic put Mythos 5 back in front of a limited set of trusted US customers. That solves the immediate outage, but it also proves something every AI buyer now has to price in: Washington can pull a frontier model from the market fast.
The letter came Friday from Commerce Secretary Howard Lutnick to Anthropic chief compute officer Tom Brown. Mythos 5, the company's most capable enterprise model, can be restored for approved US organizations and their foreign national employees, according to reporting from Axios and The Wall Street Journal. Fable 5, the consumer-facing model caught in the same June 12 order, is still blocked for general use.
That distinction matters. If you're a bank, a defense contractor, a law firm, or a federal agency that built work around Anthropic's latest model, this wasn't an abstract policy fight. It was two weeks of procurement teams asking whether a signed vendor contract means much when the government can step in with a national security order and the vendor has to obey in minutes.
The chain of events started on June 11. Amazon CEO Andy Jassy raised concerns with White House officials about a vulnerability in Fable 5, according to earlier reporting on the dispute. Officials directed him to Treasury Secretary Scott Bessent, and Jassy spoke with Bessent that day. By the next evening, Anthropic had received a Commerce Department directive restricting access to Mythos-class models. The company pulled Mythos 5 and Fable 5 worldwide because it couldn't practically enforce nationality-based restrictions across all users.
The alleged jailbreak was narrower than the panic around it suggested. Anthropic has described the government evidence as a potential non-universal exploit that involved asking a model to inspect a codebase and fix software flaws. That is not the same as a model casually breaking into classified systems on its own. But The Economist's earlier reporting, later picked up by Tom's Hardware, added a more alarming detail: Senator Mark Warner said Gen. Joshua Rudd, who leads the NSA and US Cyber Command, told him Mythos broke into almost all of the agency's classified systems within hours during an authorized red-team evaluation.
Look, those two claims don't sit neatly together. One sounds like a bounded coding task. The other sounds like a national security alarm bell. The important fact for customers is simpler: Commerce acted before the public had a clean technical explanation, and Anthropic complied.
The commercial hit was real. Anthropic had been gaining ground in enterprise AI, helped by Amazon and Google backing and by a reputation for models that large companies could put near sensitive work. Mythos 5 was meant to deepen that lead. Instead, for two weeks, OpenAI and Google could tell nervous buyers something Anthropic couldn't: our flagship model is still available.
Don't treat the Friday letter as a full reversal. Axios reported that Lutnick's June 26 letter says a license will no longer be required for Claude Mythos 5 transfers to entities listed in an annex, their foreign national employees, and Anthropic's own foreign national employees. It also says export controls remain in place for everyone else, and it doesn't change the restrictions on Fable 5. The government has opened a gate, not removed the fence.
The precedent is the story
The sharpest point here isn't that Anthropic lost access to customers for two weeks. It is that the US government applied export-control logic to a hosted software model, then used that power against a domestic company serving domestic customers. Export controls have usually meant chips, semiconductor equipment, and shipments across borders. Here, the object was model access.
If you buy frontier AI, you need to understand the new risk. Your uptime clause can promise service credits. Your security addendum can describe data handling. None of that helps much if Commerce decides a model has become a controlled national security asset overnight. Frankly, every serious enterprise buyer should now ask vendors for a written answer to one question: what happens to our workflows if your best model gets restricted by the government?
OpenAI and Google got a short-term opening from Anthropic's outage, but they shouldn't celebrate too hard. The same precedent can land on them. Axios noted that OpenAI has already complained that model-by-model government reviews aren't sustainable. That's not special pleading. It's operational reality. A market where frontier models can be cleared, blocked, or partially restored by letter is a hard place to sell reliability.
Anthropic said it is working to restore access for the approved set of providers and wants to make Fable 5 generally available again. Lutnick's letter, according to Axios, also reserved the right to change the scope of licensing requirements if circumstances change. That sentence is doing a lot of work.
The company got Mythos 5 back for some customers, and that's better than a continuing shutdown. But the lesson for you isn't that the crisis is over. The lesson is that model access now depends not only on price, benchmarks, latency, and safety testing, but on whether Washington is comfortable letting the model run. That is a new procurement category, and it won't stay confined to Anthropic.
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, and The CFTC just opened a fraud probe into Polymarket and the timing could not be worse for the prediction market boom
