When Anthropic prepared its latest model for release, one established partner was left off the list. The company declined to submit the model to Britain’s AI Safety Institute for pre-deployment testing, according to the Financial Times, which cited people familiar with the matter. The model, Claude Mythos 5.1, was made available only to vetted institutions in the United States.
The decision has caused concern inside the institute and across Whitehall, the FT reported, citing British government figures. The UK AI Safety Institute was set up precisely to run this kind of check: to examine frontier models before release and report on their risks. A major lab declining to participate weakens a system that depends on voluntary cooperation.
The institute is a young institution. Britain created it in late 2023 from its Frontier AI Taskforce, formalized around the Bletchley Park summit that gathered world leaders to discuss frontier AI risk. Its mandate was to test models that no single company could be trusted to evaluate alone. Labs including Anthropic, OpenAI, and Google DeepMind agreed at the time to give it pre-release access, a set of handshake commitments rather than legal obligations.
Anthropic declined to comment on the report. The Cabinet Office said the institute “remains closely engaged with industry partners, including Anthropic, to improve model safety,” a formulation that does not confirm whether the newest model was tested. The gap between what the government says and what the company would not say is the story.
Anthropic’s willingness to work with American authorities is not in doubt. The company signed a memorandum of understanding with the U.S. AI Safety Institute in 2024 to collaborate on testing and risk research. The difference now is one of scope and sequence: American vetted institutions got access to Claude Mythos 5.1, while Britain’s institute, its longest-standing international testing partner, did not. People familiar with the matter told the FT the omission was deliberate.
The episode marks a shift in how the frontier labs treat government testing. Anthropic was an early and willing participant in voluntary safety evaluations, and its executives have spoken often about the need for outside scrutiny of powerful models. Choosing to route Claude Mythos 5.1 only through American channels, and only to approved institutions, suggests the company is now being more selective about who sees its work before it ships.
The timing matters. Anthropic is preparing for an initial public offering, and it has leaned on a “safety first” identity as part of its valuation story. A public dispute with a national safety institute, even a quiet one, cuts against that positioning at a moment when investors are listening closely to the company’s claims about how responsibly it operates.
Anthropic’s decision also arrives as the company courts public investors. Executives have spent years describing the lab as one that would slow down rather than risk harm, and that framing is now under strain from within, including the resignation this week of a researcher who said the lab was racing toward systems no one can control. Declining to show a frontier model to Britain’s safety body adds an external data point to a narrative the company would rather not be writing.
Britain’s institute is not a regulator with the power to compel submissions. It relies on agreements with labs that have, so far, mostly cooperated. When a company opts out, the institute can do little beyond noting the absence. That voluntary structure is now being tested by a lab that appears to have concluded the benefits of an early UK review no longer outweigh whatever it costs.
The broader question is whether frontier model testing is becoming a bargaining chip. Labs compete to release the most capable models first, and a weeks-long government review can slow a launch. If one lab skips the queue, others face pressure to follow, and the testing regime that several governments have spent years assembling could erode model by model.
Britain is not the only government watching. Washington, the European Union, and other capitals have built or are building similar testing bodies, each depending on the same voluntary goodwill. If labs conclude that skipping evaluations is cheap, the institutes risk becoming bystanders to releases they were created to scrutinize, and the safety record they keep will reflect whatever companies choose to hand over.
For now, the consequence is a hole in the safety record of one of the most capable systems ever released. British officials will have to judge Claude Mythos 5.1’s risks from a distance, and the public will have to trust that a company which declined to show its work has still done it. That is a thinner basis for confidence than the institute was built to provide. For a body whose purpose is to look before a model ships, being told to look only after it ships is close to being told not to look at all.
