White House Wants First Look at New OpenAI and Anthropic Models

The White House has told OpenAI and Anthropic to hold back new AI models from the U.K.'s AI Safety Institute until U.S. government agencies get to review them first.

That single instruction, reported by The Decoder, reverses the sequencing of a cross-border testing arrangement that had become a quiet fixture of frontier-AI oversight. Under the existing setup, Britain's AI Safety Institute (AISI) received early access to new models from leading U.S. developers. Now the White House wants American reviewers in the room before anything crosses the Atlantic.

The demand applies to two of the three most valuable AI developers in the United States. OpenAI, the company behind the GPT model family, and Anthropic, maker of the Claude model line, have both made voluntary commitments to submit frontier models for pre-deployment safety testing. The U.K.'s AI Safety Institute, established under the previous British government to evaluate advanced AI systems, has been one of the main beneficiaries of those commitments.

What the directive changes

At stake is the order of operations. Under the arrangement the White House now wants to alter, new models could reach British testers before — or alongside — U.S. reviewers. The new instruction establishes the United States as the first stop: OpenAI and Anthropic are expected to let U.S. agencies review new models before sharing them with the U.K. institute.

For the companies, the practical effect is a scheduling constraint on their testing pipelines. Pre-deployment evaluation has become a standard part of releasing frontier models, and the major developers coordinate with multiple government testers. Where Britain previously sat at or near the front of that queue, the White House is asserting priority for its own agencies.

For the U.K., the move signals a downgrading of the access its safety institute enjoyed. AISI's evaluation work depends on developers granting it early access to systems before public release. If U.S. review now comes first, British testers receive models later — or, in the worst case for them, on terms set in Washington rather than London.

Why it matters

The directive lands at a moment when government access to frontier models has become a contested policy question in both countries.

In the United States, federal agencies have been building their own capacity to evaluate advanced AI systems, and the White House has repeatedly emphasized that American-built models should serve American interests. Telling the two leading U.S. developers to let domestic agencies review models first fits that posture. It also gives U.S. reviewers earlier visibility into systems that are central to the country's economic and security ambitions.

In the United Kingdom, the AI Safety Institute was positioned as a flagship institution of the country's AI policy, hosting the 2023 Bletchley Park summit and cultivating working relationships with OpenAI, Anthropic and other developers. Its testing program has relied on voluntary cooperation rather than binding regulation. The White House directive shows how fragile voluntary arrangements can be when a bigger government decides to reorder them.

The episode also illustrates the asymmetry of power in international AI governance. The United States hosts the leading model developers; Britain does not. Whatever testing regime Washington prefers will, in practice, shape what other governments can demand. A small number of companies — OpenAI and Anthropic among them — now mediate between competing national regulators, and each sequencing decision reflects a negotiation over who gets to see frontier capabilities first.

The open questions

The report leaves several things undetermined. The White House has not publicly specified which U.S. agencies would conduct the reviews, how long they would take, or whether the requirement will take any formal shape beyond a directive to the companies. OpenAI and Anthropic have not publicly confirmed how they will respond, or whether the change delays any planned releases.

It also remains unclear whether other developers that share models with the U.K. institute — Google DeepMind and Meta have made similar voluntary commitments — will face the same instruction, or whether the directive targets only OpenAI and Anthropic.

What is clear is the direction of travel. The White House has decided that when it comes to new models from OpenAI and Anthropic, U.S. reviewers come first and British testers wait — and the burden of adjusting now falls on the companies and on the U.K. institute that had grown accustomed to early access.