After a Friday US export control order, Anthropic pulled its newest public AI model, Fable 5, and the underlying Mythos model offline, becoming the first real test of the Trump administration's new AI regulation regime.
Anyone who tried to use Fable 5 in Anthropic's Claude chat this week saw a short, unceremonious notice above the message box: "Fable 5 is currently unavailable." As of The Verge's recording, the message was still there, and the model was still offline. There was no estimated restoration date. The reason for the outage was not a routine maintenance window or a guardrail tweak. It was an export-control order from the US government, and Anthropic's conclusion, in the company's own framing, that running the model under the new rules was not workable.
Fable 5 is Anthropic's newest publicly released AI model, shipped to the public less than a week before the order. The export-control notice also covered Mythos, the underlying model referenced alongside Fable 5, which is why both came down together. According to The Verge's reporting, the order restricted access for foreign nationals, including those working for Anthropic inside the United States. Anthropic concluded it could not reliably restrict access and stay compliant, so it took both models offline for everyone.
This is not a generic export-control story. It is the first concrete test of the Trump administration's new AI regulation regime, and it landed on a company that built its brand on the opposite premise: that frontier AI is dangerous enough to require unusual caution. The export-control order is the trigger, but the live question is not which model is offline. It is which institution has the authority to decide when an AI model is "too dangerous" to share, and on what evidence.
Until last week, the practical answer was the company. Anthropic, like its peers, set its own release terms and decided what was safe enough to ship. That arrangement is now visibly contestable. The Trump administration, through an export-control mechanism, can effectively pull a frontier AI product out of the market without forcing a court fight. The company responds by complying or, as in this case, by shutting the model down. There was no court order, no hearing, and no public docket. There was a list of access restrictions, and a market participant deciding it could not live with them.
That mechanism is the part of the story worth holding onto, because it reframes a one-day outage as a governance question about AI. The export-control route lets the federal government treat AI models the way it already treats other advanced technology subject to licensing requirements: as items that can be licensed, restricted, or pulled from foreign access. For chip exports, the political friction has usually been about China. For AI models, the friction here is internal: the controls reach inside US borders and apply to foreign nationals on US soil who work for an American AI company. Anthropic's compliance posture is what made that reach consequential.
The Anthropic irony is structural, not editorial. The same company that spent years arguing advanced AI might soon be dangerous, and that helped put frontier AI risk on the policy map, is now publicly objecting to the government's call that its own newest model was sensitive enough to restrict. As The Verge's Decoder podcast framed it, the question is not whether AI can be dangerous. The question is who gets to make that call when the AI in question is one of yours.
The honest answer right now is that the Trump administration can, and has. Anthropic's choice was compliance, and the cost of compliance was a model that has been offline for days. Other US AI companies with frontier products are now reading the same list of access restrictions and asking the same question: can we keep our model available, and our foreign-national engineers working on it, under these rules. If the answer is no, expect more models to come down before the regulatory text is fully public.
What to watch next: whether the underlying export-control order is published in a form that lets outside researchers evaluate the criteria, whether Anthropic formally challenges the order or quietly negotiates a narrower licensing path, and whether other frontier AI developers preemptively restrict access for foreign-national staff to avoid the same forced takedown. Fable 5's chatbox notice is currently the visible surface of a much larger argument about who, exactly, is qualified to decide when an AI model has crossed the line.