Two weeks of regulatory purgatory for Anthropic's most powerful AI model is starting—emphasis on starting—to lift. The Commerce Department has given the AI lab a conditional green light to restore access to Claude Mythos 5 for a limited set of US organizations: over 100 entities, a mix of large corporations and government agencies that the administration has deemed sufficiently trustworthy.

Commerce Secretary Howard Lutnick communicated the decision in a letter to Anthropic cofounder and chief compute officer Tom Brown, stating he had "determined that appropriate safeguards are in place." Semafor first reported the letter's existence; WIRED obtained a copy. That's about as warm and fuzzy as regulatory approval gets these days.

What Actually Happened Here

Let's back up. On June 12, the Trump administration sent Anthropic an export control directive requiring the company to cut off access to both Mythos 5 and its consumer-facing sibling, Claude Fable 5, for foreign nationals—including people legally working and living in the United States. Anthropic's response? Pull the plug on both models entirely, rather than try to surgically enforce nationality-based access controls at the user level. That's not an unreasonable call, by the way. Verifying residency status in real time at inference is its own engineering nightmare.

The trigger for all this: the White House reportedly learned that Anthropic had granted Mythos access to a South Korean telecom firm it believed had ties to China. Separately, Amazon and the NSA flagged concerns that Fable 5 could be jailbroken. When two national security heavyweights are raising red flags about the same model in the same week, bureaucratic inertia tends to get replaced by bureaucratic panic.

The partial restoration now allows approved organizations to let their foreign national employees use Mythos 5. Anthropic can also restore access for its own foreign national staff. That's a meaningful rollback of the original directive's most operationally disruptive element—you can't exactly run an AI lab if half your engineering team can't touch the flagship model.

The Part That's Still Unresolved

Here's the catch: Claude Fable 5, the consumer-accessible version with additional safety guardrails baked in, is still in limbo. Anthropic spokesperson Eduardo Maia Silva confirmed the company is "working to provision the approved set of providers and restore their access to Mythos 5 as quickly as possible" but noted that discussions about Fable 5's fate are ongoing—apparently including weekend negotiations with the White House. Nothing says "high-stakes AI policy" like burning through your Saturday to talk export controls.

Both parties, according to sources familiar with the situation, are hoping this incident becomes a template for a more durable policy framework governing future frontier model releases. Which sounds promising until you realize that "we're figuring out the rules as we go" is exactly how you end up with another emergency shutdown two product cycles from now.

The Bigger Problem Nobody Wants to Say Out Loud

Dean Ball, head of strategic futures at OpenAI and a former White House AI adviser, put it bluntly in a mid-June blog post: frontier AI developers now need "an explicit green light from the government" before releasing models. That's not a hypothetical future state—it's apparently the current operating environment. OpenAI underscored the point on Friday by announcing it was delaying its GPT-5.6 release at the administration's request.

Think about what that means structurally. You've built a model that cost hundreds of millions of dollars to train. You've run your safety evaluations. You're ready to ship. And now there's an informal (or not-so-informal) pre-clearance process sitting between you and your customers, operated by an administration that is simultaneously trying to win an AI race with China and apparently concerned enough about your specific model to shut it down on a Friday evening.

This isn't purely about Anthropic's choices or Mythos's capabilities. It's about the emergence of a de facto regulatory regime that nobody voted on and that has no formal statutory basis yet—just letters from the Commerce Secretary and a lot of weekend phone calls.

What It's Costing Anthropic

The business damage here is real. Anthropic already sued the Trump administration earlier this year over a supply chain risk designation it received after trying to set limits on how military contractors could deploy its models. That's a company that has now both litigated against its largest potential government customers and had its flagship products switched off by executive directive. The reputational math on enterprise sales is not great when your product could disappear from your customer's workflow with a Friday-evening email.

In response, Anthropic sent senior members of its cybersecurity and AI safety teams to Washington for face-to-face meetings with administration officials. Brown and public policy chief Sarah Heck have been leading the Commerce Department negotiations. That's a significant deployment of senior talent on regulatory firefighting rather than, say, building the next model.

The partial reinstatement of Mythos 5 is a genuine step forward—Anthropic's cybersecurity-focused customers can breathe again, and the foreign national access issue is at least partially resolved. But until Fable 5 is back online and there's some kind of coherent framework for what "government approval" actually means in practice, every frontier AI lab in the country is now operating under the shadow of a kill switch they don't control. That's a tradeoff worth watching very carefully.