What METR Concluded

On November 5, 2026, METR published a public report on its August 12 to October 29 assessment at Anthropic. Most METR staff on the project judge that Anthropic, as of October 2026, was likely advancing its models on well-specified tasks about twice as fast as between March 2024 and October 2025. The judgment is subjective and highly uncertain: a figure of 1.3 or 3 times would not surprise METR, and broader tasks are less certain.

The main public measurement is AECI, Anthropic’s adaptation of the Epoch Capabilities Index, which includes unreleased internal models. Since October 2025, AECI has grown about 1.9 times as fast as before (95 percent confidence interval 1.5 to 2.8), and alternative specifications give 1.1 to 2.9. This trend was a major input to the judgment, though benchmark overfitting or the AECI method could partly explain it.

METR roughly estimates that perhaps 80 percent of the acceleration comes from AI use, with substantial uncertainty. It saw convincing evidence that faster training compute growth did not cause the acceleration; the details are redacted. METR understands that AI cannot yet make major, fully autonomous contributions to core model training, though researchers report early signs of them.

Redaction and Access

Anthropic, which held unilateral redaction rights, removed most of the evidence METR found most informative, METR writes. Anthropic says the redactions protect confidential information that could speed up other AI developers. METR says the public must therefore rely on its judgment far more than is ideal. The roughly 100-page unredacted Full Report went to Anthropic on October 22 and is not public. Midway through, Anthropic also cancelled a planned check of how it allocates compute, including between capabilities and safety.

METR saw non-public datasets and documents that Anthropic approved but had no direct system access or queries to unreleased models. METR says the project never aimed at full employee-like access, which would mean unmediated use of data, documents, and models, and is separate from the arrangement Anthropic announced on October 15.

Audit Implications

Our editorial reading: the twofold figure is an evaluator’s judgment about capability progress, informed in part by AECI. It is not a measured doubling of researcher productivity or evidence of autonomous recursive self-improvement. When the developer controls redaction, high access alone does not make conclusions publicly verifiable. Open questions: will another qualified reviewer see the Full Report, and will the compute check resume?