The export control that is not in the licence

The story circulating since Tuesday is that Anthropic has written an export control into its newest model: the company asked Washington for international rules on frontier AI, was refused, and imposed a version of one on itself. The underlying fact is real. The mechanism is not what the word licence implies, and the difference turns out to be the substance.

Here is what the company has actually published. A support article explaining why Claude sometimes changes model mid-conversation lists four categories of automated classifier: offensive cybersecurity, biology, distillation, and a fourth that exists only on the newest model. "Opus 5.5 has classifiers similar to Fable models for a small set of capabilities related to the development of frontier LLMs, such as kernel development for certain ML accelerators," the article says, adding that this "shouldn't impact the vast majority of traditional AI or ML development". When it fires, Claude "re-runs your blocked request on a less capable model in the same conversation" — here Opus 5, the model Opus 5.5 replaced at the top of the line on Tuesday.

Which accelerators? Anthropic has not said. The phrase is "certain ML accelerators" and there is no list anywhere in its documentation. The identification comes from outside. On 22 September, the day Opus 5.5 shipped, the independent researcher who posts as xlr8harder reported that prompts about kernel development for Huawei's Ascend 950DT triggered the fallback, and that so did prompts about Amazon's Trainium3. WCCFTech reported those results the following day. Neither company has commented, and nobody outside Anthropic can say whether the Amazon result is deliberate or an artefact of how the classifier generalises.

That possibility is worth sitting with, because Amazon is Anthropic's principal compute partner and Trainium is the silicon Anthropic trains and serves on. A restriction meant to stop an American model accelerating a Chinese accelerator programme appears, on the available testing, to apply equally to the hardware of the company paying for the data centres. If that is intentional, it is striking even-handedness. If not, it shows exactly what kind of instrument this is.

An export control and a classifier are different objects. An export control is aimed at a person: a national, an entity, a destination. It is published, it creates a legal duty, and breaching it is an offence. A classifier is aimed at a question. It cannot tell who is asking, so it catches a partner and a rival with the same indifference, and nobody is in breach for having asked.

Anthropic does operate a genuine jurisdictional restriction, worth distinguishing from this one. The Usage Policy in force since 15 September 2025 prohibits using inputs and outputs "to train an AI model (e.g., 'model scraping' or 'model distillation') without prior authorization", and prohibits facilitating access "to persons, entities, or users in violation of our Supported Regions Policy". That is where the geography lives: mainland China, Hong Kong and Macau are absent from the supported list, and the company reserves the right to refuse entities majority-owned from nations not on it. Those terms are contractual, enforceable against a customer, and long predate this week.

What they do not contain is a single word about chips. Neither does the launch announcement for Opus 5.5, which describes the model's alignment testing at length and its accelerator classifier not at all. The only place the restriction is written down is a help-centre page about why the model picker changed.

The consequences are not trivial. In the consumer product the effect is silent in the way that matters: the user is told the model switched, not what tripped it, and cannot contest a false positive. In the API the default differs again. "Automatic switching isn't active by default," the support article says; customers must configure fallbacks, and until they do "the model will return a 200 response with a stop reason". The restriction is therefore least visible, and most easily retried around, among exactly the buyers most capable of frontier kernel work.

Set against the week's politics, the move reads as something other than defiance. On Tuesday in New York the heads of OpenAI, Anthropic and Hugging Face used a United Nations platform to ask for international coordination. Sam Altman said that "if AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone". Michael Kratsios, the president's technology adviser, replied that "international dialogue in this forum and others cannot be allowed to drift toward global governance", the BBC reported. Washington and Beijing had already rejected the pacing proposal inside 48 hours earlier this month.

So a laboratory twice told there will be no public rule has installed a private one. It required nobody's agreement, took effect on a Tuesday, and applied worldwide from the moment the weights went live. That is the trade on offer in the absence of a standard: speed and reach, in exchange for a restriction undocumented in its scope, unappealable in its operation and revocable without notice.

Three things would clarify matters. Anthropic could publish the list of accelerators, which would turn a rumour into a policy. It could say whether Trainium3 belongs on it. And it could put the restriction where a compliance officer would look. Until then, the most accurate description of what shipped on Tuesday is not an export control. It is a company declining to answer certain questions, and saying so in the help centre.