The Trump administration's export ban on Anthropic's Fable 5 and Mythos 5 models is still in force two weeks after it landed, and the negotiated exit is starting to look less like a narrow fix and more like a permanent change in who controls frontier AI.
On June 12, at 5:21 p.m. ET, Commerce Secretary Howard Lutnick sent a directive to Anthropic ordering the company to suspend all access to its two most powerful models for any foreign national, anywhere on earth, including non-citizen employees sitting at desks inside Anthropic's own offices. Because Anthropic can't verify citizenship in real time across hundreds of millions of users, it did the only thing it could: it pulled both models for everyone. As of June 25, Anthropic staff confirmed the company is serving exactly zero Fable 5 or Mythos 5 traffic.
The stated trigger was a jailbreak. According to reporting by Axios and later confirmed by the company itself, the government became aware of a method for circumventing Fable 5's safety guardrails and gave Anthropic verbal notice of the vulnerability before issuing the ban. Anthropic disputed the proportionality of the response, saying it had received notice of "a potential narrow, non-universal jailbreak" and disagreed that a single exploitable edge case should be grounds for pulling a commercial model deployed globally. What makes the conditions for reinstatement more striking than the ban itself is what Lutnick's June 16 follow-up letter demanded: not just a fix to Fable 5, but a guarantee that no guardrails across any of Anthropic's frontier models can be circumvented, plus ongoing jailbreak testing with mandatory government notification of any issues found. That's not a product recall. That's a compliance regime.
Over 50 cybersecurity executives, including senior figures from Nvidia and Adobe and former Facebook Chief Security Officer Alex Stamos, signed a letter urging the administration to reverse course. Their argument wasn't that Fable 5 is perfectly safe. It was that pulling America's best cyber defense tool while China's models are "only months behind the best American ones" hands the advantage to adversaries without meaningfully reducing risk. The letter asked the government to "commit to an open, scientific and transparent process of handling AI risk assessments in the future" , a pointed way of saying the current process isn't any of those things.
On June 23, Legion LegalTech Corp filed suit in Washington, D.C. federal court, arguing the Commerce Department exceeded its statutory authority because no existing export control law actually covers hosted AI models or their outputs. The case may become the first major court ruling on whether export control statutes written for physical goods and software can be stretched to cover API access to a language model. That's a genuinely open legal question, and the answer will matter far beyond this specific ban.
Meanwhile, the Globe and Mail and Bloomberg have both reported that Anthropic and administration officials are working toward a deal to restore access. The shape of that deal, if it happens, will say more about Washington's long-term intentions than the ban itself did.
OpenAI is watching this very carefully
The reason this matters beyond Anthropic is what happened at OpenAI in the same week. On June 26, as reported by CNBC, OpenAI announced three new models (GPT-5.6 Sol, Terra, and Luna) but complied with a White House request to limit their rollout to a "small group of trusted partners" initially. OpenAI stated plainly that it doesn't believe "this kind of government access process should become the long-term default," but it cooperated anyway. Trump signed an executive order earlier this month asking AI companies to voluntarily submit advanced models for government review 30 days before release, though the framework for those assessments hasn't been built yet.
So within the span of two weeks, the two most commercially significant AI labs in the United States both launched frontier models and both had the government intervene in how those models could reach users. One got a ban, the other got a soft throttle. Neither got to ship freely. That pattern, if it holds, is the actual story here, not the jailbreak that nominally triggered the Anthropic ban.
Frankly, the Lutnick conditions point toward something like a pre-clearance regime, where labs must demonstrate safety compliance to a government standard before a model can be made broadly available. The administration hasn't formalized that framework yet, but it's already acting as if one exists. OpenAI is actively working with the Trump administration to establish what it calls "a repeatable process for future model releases." Repeatable processes have a way of becoming permanent ones.
For Anthropic, the immediate question is whether it can negotiate terms it can actually operate under without setting a precedent that makes the next ban easier to impose. For the rest of the industry, the question is whether the government's current behavior represents emergency improvisation around a genuinely alarming model, or the opening move in a broader assertion of control over what AI the world gets to use. Right now, both things could be true at once.
Also read: Regulators are finally building the AI they need to police the markets they oversee • Meta is building a prediction market app called Arena and the existing players should be worried • The New York Times is now arguing Microsoft built a machine specifically designed to steal journalism