On June 12, the federal government did something it had never done before. It ordered a company to switch off an artificial intelligence model. The company was Anthropic. The targets were its two most powerful models, Fable 5 and Mythos 5. The order arrived by letter from the Commerce Department's export-control office, gave Anthropic ninety minutes to comply, threatened criminal and civil penalties, and demanded that the models be blocked for every "foreign national" on Earth, including Anthropic's own employees who are not U.S. citizens. No company can verify the nationality of every user in real time, so Anthropic did the only thing it could and turned the models off for everyone. Within hours, hundreds of millions of people lost access.
One of them was a small company in San Jose called Legion. Legion builds software that helps lawyers draft documents and manage their cases, and that software runs on Anthropic's AI. Several of Legion's engineers are Canadians working from Canada, and they were using Fable 5 to build the product when access vanished. On June 23, Legion, represented by the firm Spencer Fane, sued the federal government in Washington's federal district court, asking a judge to declare the shutdown unlawful and have it reversed. The case is Legion LegalTech Corp. v. United States. Anthropic itself is not part of the case.
Strip away the legal vocabulary and Legion's claim is straightforward: the government did something the law never gave it the power to do.
The catch that shapes the whole lawsuit
That claim runs into an early obstacle, and it explains why the lawsuit is built the way it is. Normally, when the government issues an order you think is wrong, you challenge it under a law called the Administrative Procedure Act, which lets courts review whether an agency acted reasonably. But the rules governing export controls, which limit sending sensitive technology to other countries, specifically block that kind of review for these decisions. So Legion leans on an older, narrower type of lawsuit, one that argues an official didn't just make a bad call but acted entirely outside the law. Courts rarely agree to that. To win, Legion has to show the overreach is obvious, not merely arguable.
It makes three runs at that.
Argument one: the rules don't reach it
The first is that export-control law simply doesn't reach what happened. Those rules work off an official government list of controlled items, each with its own code. Back in early 2025, the previous administration had added a code to that list, known as 4E091, that covered the "model weights" of the most advanced AI systems. Model weights are the trained numerical core of a model, the part that actually does the thinking, and they were the one piece of AI ever placed under direct export control. But in May 2025 the current administration moved to scrap that rule and told its own officials not to enforce it, promising a replacement that, by Legion's account, never arrived. So Legion's argument is that the only control that ever reached AI model weights is effectively gone, and you cannot break a rule that is no longer on the books. And even if it still existed, Legion says, it would not matter here, because that rule was about handing over the model weights themselves. When you type a question into a chatbot and it sends back text, none of that core technology changes hands. You get words, not the machinery that produced them. The rules even spell out that finished program code does not count. Legion adds that the two specific legal hooks the government cited do not fit either: one requires a public rule-making process that never took place, and the other is a narrow tool meant for specific military buyers in a handful of named countries like China and Russia, not a blanket worldwide ban.
Argument two: you can't ban the flow of information
The second argument is a backup, in case the government says it was using emergency presidential powers instead, the same kind presidents use to impose sanctions. Here Legion has a clean point. A law from the 1980s, known as the Berman Amendment, says those emergency powers cannot be used to block the flow of information, things like books, films, and news. Legion argues that the legal drafts and written analysis its AI produces are exactly that kind of information. On top of that, emergency powers require the President to formally declare a national emergency over a threat that comes from outside the United States. No emergency was declared, and the supposed threat here is a product built by an American company and hosted in America. Legion also says that effectively forcing every powerful AI model to get government approval is an enormous new power, and the Supreme Court has held that sweeping powers like that need clear permission from Congress, which Legion says is missing.
Argument three: it makes no sense on its own terms
The third argument is the most intuitive: even on its own terms, the decision doesn't make sense. The stated problem was narrow. Someone found a way to get the model to examine software and spot flaws in it, something most chatbots can do. The response was to cut off the entire world. Worse, Legion says, it accomplishes nothing, because the same code-analyzing ability is freely available in rival products like OpenAI's GPT-5.5, which the order left untouched. The government, Legion claims, ignored milder options such as blocking only people from hostile countries, ignored the businesses that had already built products on these models, and didn't do its own homework. It allegedly acted on a report and a few phone calls from Anthropic's competitors over a single evening.
The feud underneath
Hanging over all of this is a feud. According to the complaint, the administration had been clashing with Anthropic for months after the company refused to give the Pentagon unrestricted access to its tools, insisting only that they not be used for mass surveillance or autonomous weapons. The filing says the President ordered agencies to stop using Anthropic, that the Defense Secretary branded it a "supply chain risk," and that the President warned of "major civil and criminal consequences." It notes that a federal judge in California reportedly already found an earlier national-security move against Anthropic was likely retaliation. And it points out that just ten days before the shutdown, the President signed an order promising the government would not require licenses or pre-approval for AI models. Then, Legion says, it did precisely that. The thrust is that "national security" was a cover story for punishing a company that wouldn't fall in line. These are allegations, not proven facts, and the government has not yet responded.
What the government will say back
The government will have answers. The kind of lawsuit Legion filed is one courts almost never allow, and judges give the executive branch wide latitude on anything labeled national security. There is also a practical snag: even if a court throws out the order, access only returns if Anthropic chooses to restore it, and Anthropic isn't in the case. Legion's claim of harm also leans partly on what would happen over time, the lost ground and the threat to its survival, and the government may argue those forward-looking losses are too speculative to justify emergency relief. And the information-flow argument, while clever, is untested. The government will say it regulated an interactive service, not a published work.
Why it comes down to speed
The real fight won't be over whether the case can survive. It will be over speed. Legion is asking for an emergency order to turn the models back on while the lawsuit plays out, and to get it the company has to convince a judge it is likely to win and is suffering harm that money cannot repair later. Its pitch is that in a field moving this fast, time lost is gone for good, a small team loses the edge that lets it compete with giants, and because you cannot collect damages from the government, only a court order can help.
Why it matters
Whatever the outcome, this is the first time a court will decide whether the government can reach in and switch off a commercial AI model that millions of people already rely on. If the government wins, it will have established that it can disable these tools more or less at will. If Legion wins, that power will have a limit. For anyone building a business on top of someone else's AI, and that is now a very large group, it is hard to think of a question that matters more.
The complaint described in this article is a pending matter and reflects allegations and arguments that have not been finally adjudicated. The government had not responded as of publication, and case status is current as of June 24, 2026. The consumer-facing tracker of the AI docket at Lawsuit Informer is updated as the cases move.