Anthropic opens models to outside safety auditors as Amodei calls for slowing AI frontier

Anthropic chief executive Dario Amodei is calling for the AI industry to ease off the accelerator, and his own company is going first. In a lengthy essay, Amodei laid out a three-step proposal he calls "pacing the frontier" — essentially a framework for slowing how quickly the most capable AI systems get built and released — and committed Anthropic to giving outside evaluators access to its models. Among those evaluators is METR, the nonprofit that runs independent assessments of frontier systems.

The concrete part of the announcement is the access arrangement. Anthropic says third-party groups will be able to examine its models to verify that the company is living up to its stated safety practices and commitments, rather than relying on Anthropic's own internal reviews. METR, formerly the Center for AI Safety's evaluations arm, has previously tested models from OpenAI and others, and its reports have become one of the few public windows into how frontier systems behave under stress. Handing it keys to Anthropic's models gives those assessments more reach and puts a named auditor behind the company's safety claims.

Amodei's argument rests on a timing problem rather than a technical one. If labs keep shipping more capable models on the same aggressive schedule, he contends, safety work and outside scrutiny cannot keep up, and the competitive pressure to release first makes the problem worse. Slowing the frontier, in his framing, is not a call to stop development but to insert checkpoints — evaluations, disclosures, and possibly coordinated pauses — before the next tier of capability goes out the door. That is a notable position for the head of a company that is itself racing to build frontier models.

The proposal lands in a crowded policy moment. Governments in the US, UK and EU have spent the past two years building evaluation bodies and voluntary commitments, while labs have signed agreements that critics call light on enforcement. Anthropic's move gives regulators and researchers a test case: whether a leading lab will actually accept outside judgment when the findings are unflattering. It also raises the question of whether rivals will follow or treat the slower approach as a competitive opening.

Amodei did not attach a timeline or say which future models would be covered, and the essay stops short of naming specific capabilities that would trigger a pause. That leaves the plan's teeth unclear. The near-term test is whether METR and other evaluators publish findings on Anthropic's systems with enough detail to matter — and whether Anthropic keeps the arrangement in place if the results complicate its next launch.