Anthropic grants METR deeper access for AI safety reviews

The strategic issue here is who can see enough inside a frontier lab to challenge its conclusions. Anthropic says METR's embedded team receives access roughly comparable to its internal risk assessors, with publication rights that the company cannot veto outright. Those terms make this more consequential than an external review limited to a finished model's public interface. ([anthropic.com](https://www.anthropic.com/news/an-update-on-independent-evaluators))

Our assessment is that this arrangement could make safety evidence a more meaningful competitive differentiator. If evaluators can explain failures and disputed redactions publicly, customers and policymakers gain another basis for comparing labs. The important test will be the substance of those reports: whether they identify overlooked risks, change deployment decisions and distinguish reassuring claims from demonstrated controls. Access alone does not establish that any of those outcomes will follow.

For the AGI race, the effect on timing remains uncertain. Better scrutiny could expose reasons to delay a release, or help resolve uncertainty that otherwise blocks deployment. Readers should treat the announcement as a change in oversight infrastructure, not evidence that models have become safer or more capable.