On June 12, 2026, the U.S. Commerce Department sent a letter to Anthropic ordering the company to immediately block all foreign nationals from accessing its two most advanced AI models. Within hours, Anthropic disabled Claude Fable 5 and Mythos 5 entirely, for everyone, because it had no way to verify the nationality of every user in real time. The models came back online on June 30, but in a significantly weakened state. This sequence of events, compressed into less than three weeks, raised questions that will shape AI governance for years.
A Law Written for Hardware, Applied to Software
The Commerce Department issued its directive under the Export Control Reform Act of 2018, a statute originally designed to regulate the movement of physical goods with national security implications, such as uranium enrichment centrifuges. This was the first time the government had applied export controls to an AI service.
The legal stretch involved is substantial. When a user sends a prompt to an AI model hosted on U.S. servers and receives a response, the model itself never moves. Only the output travels. The Commerce Department’s own past guidance had treated remote access to software running on U.S. servers as outside the scope of export controls. The fact that Congress is now considering legislation to change this situation suggests that existing law may not clearly cover AI model outputs at all.
A second legal question concerns procedure. The mechanism used in the Anthropic letter, known as the “is informed” process, is typically used to notify a company that a specific type of transaction with a specific country requires government approval. The Anthropic directive went considerably further: it applied to all foreign nationals, anywhere on the planet. Even if an AI response were legally classified as an export, the global scope of the order may exceed the authority the law actually grants to the department.
For comparison, the government has a standard temporary classification process, designated by the code 0Y521, for situations requiring rapid action. That process requires sign-off from the Defense and State departments, gets published publicly, expires after one year unless renewed, and commits the government to coordinating with allies. The Anthropic letter had none of those features. It was unilateral, secret, open-ended, and global in reach.
The Models Came Back Weaker, and That Is the Point
Anthropic complied with the order and then negotiated its removal. The Commerce Department lifted the restrictions on June 30 after Anthropic strengthened the safety guardrails on both models. The outcome, however, was not a clean resolution. The new guardrails caused a measurable collapse in the models’ benchmark scores. According to one experiment described in the source material, the updated Fable 5 completed only 3 of 12 tasks that would have been routine before the new controls were applied. Many ordinary, benign queries now trigger the guardrails.
The government, in other words, agreed to restore access only after the models had been substantially hobbled. This is not a minor technical footnote. It means the regulatory intervention did not simply restrict who could use the technology; it changed what the technology could do for everyone.
Anthropic did not contest the order in court. The company described the episode as “a misunderstanding” and sent officials to Washington to negotiate rather than litigate. This choice was likely strategic. Just two days before the letter arrived, Anthropic’s CEO Dario Amodei had published an essay arguing that governments should have the power to block or deter deployment of frontier AI models considered too dangerous. Taking the government to court to deny that same power would have directly contradicted the company’s own stated position. Anthropic had previously sued the Trump administration over a separate matter, so the decision to negotiate here appears to reflect calculation, not simply compliance.
There is also a structural reason these issues may never reach a courtroom. The 2018 statute limits federal courts’ ability to strike down such decisions as arbitrary. Any legal challenge must demonstrate not that the order was unreasonable, but that it was flatly unauthorized or unconstitutional. That is a significantly narrower legal path.
Why This Matters Beyond One Company and One Letter
Here is what most coverage of this episode misses: the Anthropic case is not primarily about Anthropic. It is about the emerging architecture of state control over AI capabilities.
The incident demonstrates that a government agency can, with a single letter and almost no notice, effectively disable access to frontier AI models for a global user base. It also demonstrates that the legal framework for doing so is, at best, uncertain. The Export Control Reform Act was not written with AI in mind. The procedural safeguards that normally accompany emergency technology restrictions were bypassed entirely.
João Marinotti, an Associate Professor of Law at Indiana University who studies technology law and policy, frames the stakes clearly in the source material: the coming legal and political fight will not be about whether governments can exert this level of control, because the Anthropic case shows they already can. The fight will be about how they wield that control responsibly.
The worst-case trajectory is a world in which governments condition access to the most powerful AI systems on secret arrangements, creating a two-tiered order where the most capable frontier models are available only to those who have negotiated the right terms with the right authorities, outside any public process.
In Short
A single government letter temporarily shut down two of the world’s most advanced AI models and, when access was restored, those models performed significantly worse than before. The legal authority behind that letter is genuinely uncertain: the statute used was designed for physical goods, the procedures normally required for emergency restrictions were not followed, and the global scope of the order may exceed what the law permits. Anthropic chose negotiation over litigation, partly because its own leadership had publicly endorsed government oversight of frontier AI. The deeper issue is not whether this particular episode was handled well or badly. It is that the infrastructure for state control over AI now exists and is being used, while the legal and democratic frameworks needed to govern that control have not yet been built.
Based on reporting from The Conversation - Technology.