AI
METR estimates Anthropic’s AI-assisted research progress has roughly doubled
An 11-week METR assessment estimates that Anthropic’s model capabilities on well-specified tasks improved about twice as fast after October 2025 as during the prior baseline. Its main benchmark showed a 1.9-fold acceleration, while METR tentatively attributed about 80% of the change to AI-assisted research rather than faster compute. The conclusion remains uncertain: Anthropic redacted much of the underlying report, interviews were incomplete, a compute-allocation audit was canceled, and benchmark gains may not transfer to messy production work.
Editorial summary · Publisher article
What matters
- AECI growth rose from about 15 points per year to 29 points per year, with a 95% confidence interval of 1.5× to 2.8×.
- METR’s attribution to AI accounts for roughly 80% of the estimated acceleration but remains observational.
- Employee self-reports were much larger than the interview evidence and cannot be treated as company-wide productivity measures.
- The assessment found no clear 12-month bottleneck in compute, data, or algorithms, but that is not a firm forecast.
Evidence and uncertainty
METR combined benchmark trends, internal logs, surveys, compute information, and researcher interviews. The evidence points toward faster AI development, but redactions, saturation in the benchmark, incomplete interviews, mediated access, and confounding organizational changes limit independent scrutiny.
Context & caveats
- The report cannot establish that public releases will accelerate at the same rate or that benchmark improvements will carry over to complex research and production tasks.
Why this made the edition
The report supplies quantitative estimates, uncertainty ranges, interview evidence, methodological limitations, and clear implications, while prominently distinguishing measured findings from tentative attribution.
Source material