Anthropic keeps its newest model inside the US while Washington asks for first review

Anthropic's own announcement is about customers as well as safety testers. Claude Mythos 5.1 is "only available to a set of US organizations," the company said ^[4], and its blog tied everyone else to Washington: "We're coordinating with the US government to expand access to a broader set of domestic and international partners as quickly as possible" ^[5]. Domestic partners are in that queue too ^[5].

The US rationale is defensive. The request came from the Office of the National Cyber Director, according to POLITICO, which cited a person familiar with the matter and a senior administration official, both granted anonymity ^[1]. The official said the White House wants to review the models and secure US systems before the models reach US partners ^[2]. A review of that kind ends when the government is satisfied with its own defences. The request comes as the administration works out how to handle models that hacked into outside entities during testing, including an Australian government website ^[8].

The strongest case that little has changed comes from Henry de Zoete, director of the UK AI Security Institute. In a letter to a parliamentary committee earlier this month, he acknowledged the missing Anthropic access but pointed out that AISI tested OpenAI's GPT-6 Astra before release ^[6]. "We maintain strong relationships with all frontier AI developers and continue to have prerelease access to some of the world's most capable models," he wrote ^[6]. The US side describes something broader. "Because they're American companies and this has been our policy with every new frontier model that comes out," the senior official said ^[3]. OpenAI did not respond to requests for comment ^[9].

Britain has more at stake than either lab. AISI, created in 2023, tests the labs' latest models before public release on a voluntary basis ^[10]. The government has made that access central to its pitch to lead work on a global AI agreement ^[11]. Because the arrangement is voluntary, the labs carry the choice: withhold models from AISI, or risk upsetting the White House, as Business Insider described it ^[14]. Prime Minister Andy Burnham told the UN General Assembly the institute was "working hand in glove with the US and the leading labs on AI safety" ^[12]. "The leading AI companies have actually committed to provide this visibility. It's really, really important, and it is an offer that we and they should follow through," Foreign Secretary Ed Miliband said at the Security Council ^[13].

For an organisation building on these models, the near-term effect and the later one differ. This quarter, a team outside the US that wants Mythos 5.1 waits on coordination between Anthropic and the US government ^[5]. Next quarter, if the official's "every new frontier model" holds ^[3], a US review step sits between each frontier release and wide availability. A vendor's date for a non-US deployment then becomes a forecast of a government review. In my view, a buyer abroad is choosing between the newest model and a date it can plan around, and for Mythos 5.1 it cannot yet have both ^[4][5]. Anthropic appears to have agreed to the request so far, Business Insider reported ^[4]. Whether OpenAI does the same determines whether the US-first step governs one lab's releases or both.