METR’s November 5, 2026 assessment estimates that Anthropic’s benchmark progress on its AECI index accelerated 1.9 times after October 2025, with a 95% confidence interval of 1.5–2.8. That measures a change in the pace of progress; it does not mean a model became 1.9 times as capable.
METR judges AI responsible for most of the current acceleration, with substantial uncertainty. Staff accounts pointed primarily to AI assistance. The assessment therefore leaves autonomous research and the causal contribution of AI as separate questions. Its benchmarks also emphasize well-specified tasks, limiting conclusions about broader real-world performance.
The reported interval describes the benchmark trend estimate; causal attribution remains a judgment. METR did not sample non-public models, and its access depended on information Anthropic agreed to share.
Access changes what readers can check
Anthropic redacted much of the evidence METR considered most informative. METR says this prevents readers from fully judging the evidence themselves. This engagement also differed from the employee-like access arrangement announced in October.
In its October 15 announcement, Anthropic described onboarding six METR staff with access broadly comparable to its own risk-assessment teams. It also retained authority over which material required redaction. Readers should check which engagement a published finding actually covers before assuming that an evaluator could inspect everything.
METR’s published engagement policy makes a useful distinction between explaining the assessment process and publishing enough underlying evidence for other experts to reach their own conclusions. It acknowledges that voluntary evaluations depend on developers granting access and that some supplied information cannot be independently verified. Those constraints belong alongside the headline result when judging an assessment.
A useful reading test
For security teams, this suggests three questions when a new AI capability claim appears. What observable outcome improved? What evidence links that improvement to AI assistance? What can another investigator reproduce or challenge?
Applied consistently, those questions help teams plan evaluations without turning a single benchmark trend into a forecast of every capability. Record the assessment date and access conditions alongside the result, and revisit deployment assumptions when new evidence appears. This is a reading framework, not an independent replication of METR’s work.
