A meeting, and then nothing
Dario Amodei walked into rooms full of Trump administration officials this week to make a simple ask: give Anthropic back its two most powerful models. He walked out without it. After a round of negotiations aimed at restoring foreign access to Fable 5 and Mythos 5, no agreement was reached, no restoration timeline was confirmed, and the export control that froze both models worldwide remained in force.
The standoff is the clearest sign yet that the relationship between America's self-styled safety lab and the administration setting AI policy has curdled into something closer to a trade dispute. Anthropic is no longer arguing about a model. It is negotiating the terms of its own access to its own products, with a government that has decided frontier AI can be treated like a controlled export.
How a model became contraband
On June 12, the Commerce Department's Bureau of Industry and Security issued an export control directive ordering Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, whether located abroad or inside the United States, including the company's own foreign-national employees. Because Anthropic says it cannot reliably screen users by nationality, it disabled both models entirely for every customer on Friday, June 13.
The stated trigger was a jailbreak. The administration says a trusted partner identified a way to bypass Mythos's guardrails and that Amodei was asked to patch it or pull the model. David Sacks, the administration's AI point man, laid out that account on X, writing that the government issued the control "reluctantly" after Anthropic refused, that it wants the restriction lifted once the flaw is fixed, and that "the ball is in Anthropic's court."
Anthropic disputes nearly all of it. The company characterizes the exploit as a "narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws" — behavior, it argues, that is widely available from other deployed frontier systems. Its sharpest point is a comparison: the same exploit, Anthropic contends, applies to OpenAI's GPT-5.5, which faces no equivalent restriction. At least one security researcher who reviewed the underlying work has called it defensive research rather than a weaponized jailbreak.
That asymmetry is the heart of the standoff. If a vulnerability shared across the industry results in only one company's models being pulled, then the order looks less like a safety measure and more like a verdict on Anthropic specifically.
Strained in Washington
The export order did not arrive in a vacuum. Earlier in 2026, the Department of War designated Anthropic a supply-chain risk, a label that effectively poisons a company's standing with every federal procurement office. Negotiations with the Pentagon had already collapsed over Anthropic's refusal to drop guardrails against mass domestic surveillance and fully autonomous weapons; when the company missed a late-February deadline to accept the government's terms, agencies were ordered to stop using its products.
The pattern is consistent: Anthropic positions itself as the cautious lab, and the administration reads that caution as obstruction. The export control extended the fight from federal procurement into the open market.
This week's talks were meant to break the cycle. They covered safety protocols that might satisfy national-security concerns, frameworks for how Anthropic's models are distributed internationally, and the terms under which the export restriction might be lifted. They did not produce a deal.
The seat across the room
If the negotiating table was inconclusive, the lunch table was not. At the G7 summit in Évian-les-Bains, France, from June 15 to 17, the AI industry's three biggest names were all present. OpenAI's Sam Altman was seated to President Trump's right. Demis Hassabis of Google DeepMind was nearby. Amodei was placed across the dining room, beside Emmanuel Macron and Salesforce's Marc Benioff.
Seating charts are not policy, but in a room engineered to signal proximity to power, distance is a message. Amodei used his time to pitch a U.S.-led democratic coalition that would coordinate frontier-AI trade and standards with China deliberately outside the tent, urging leaders to "resist the temptation to splinter" their approaches. Trump, for his part, softened afterward, telling reporters that Anthropic had "behaved very responsibly" and no longer looked like a national security threat.
Warm words, but no lifted order. The gap between Trump's public thaw and Commerce's still-active directive is exactly the ambiguity Anthropic now has to navigate.
Why it matters
Two things are being tested here. The first is whether a U.S. frontier model can be reclassified, on national-security grounds, as something the government can switch off worldwide. The June 12 order is a precedent: a domestic AI product pulled from the global market by directive, over a vulnerability the maker says is industry-wide. Every lab is watching to see whether that becomes a tool of routine leverage.
The second is timing. Anthropic confidentially filed its S-1 on June 1, targeting an October Nasdaq listing at a valuation reported between roughly $900 billion and $965 billion — potentially the largest tech IPO since Facebook in 2012. An unresolved export ban, a supply-chain-risk designation, and a publicly chilly relationship with the people writing AI policy are not the backdrop a company wants in its risk-factors section. Each week the standoff drags on is a week of regulatory overhang priced into a near-trillion-dollar offering.
What to watch
Three signals will tell the story from here. First, whether Commerce formally lifts or narrows the directive, and on what conditions — a patched model, an audit, a federal-access concession. Second, whether the administration applies the same jailbreak standard to GPT-5.5 or any rival, which would test whether June 12 was about safety or about Anthropic. Third, whether the IPO timeline holds; a slip from October would be the market's verdict on how much the Washington fight is costing. For now, the models stay dark, the order stays live, and the ball, as Sacks put it, stays in Anthropic's court.
“The ball is in Anthropic's court.”— David Sacks, Trump administration AI adviser