METR: AI at Anthropic is accelerating its own research, progress runs 1.9× faster

METR spent eleven weeks inside Anthropic, from 12 August to 29 October, and published the assessment on 5 November. The team combined internal benchmark scores, code and pull-request data, agent spending, surveys and about fifteen interviews.

The headline number: on AECI, Anthropic's version of the Epoch Capabilities Index, progress accelerated 1.9× after October 2025 (95% confidence interval 1.5 to 2.8). That is roughly 15 points per year before and 29 points per year after. METR describes it as about two years of historical progress packed into one.

The acceleration was not explained by faster scaling of training compute. METR estimates that about 80% of it comes from AI, with significant uncertainty. Self-reported productivity gains rose from roughly 1.2× with Claude 3.7 Sonnet to about 12× with Model 4.

The report is unusually candid about its limits. Anthropic redacted most of the evidence METR considered most informative. Three AEF-1 checklist items, direct access, methodological disclosure and redaction authority, remain unfulfilled, and a planned verification of compute allocation was cancelled by Anthropic.

Robustness checks produced a range from 1.1× to 2.9×. METR says it would be surprised by no acceleration or by more than 3×, but calls the estimate highly uncertain. It also found no clear bottlenecks in algorithmic progress, compute or data for the next twelve months, again with low confidence.

If AI really is speeding up AI research, capability cycles get shorter. Procurement, security review and employee training then need to run on months, not years. The same logic applies to regulation: rules written for annual model releases will lag behind.

It also sharpens the debate around Anthropic's planned IPO. Investors are not buying a chatbot company, they are pricing a research engine that may be compounding faster than the market assumes. The redactions make that harder to verify.