Anthropic’s AI models have been improving on benchmark tasks at roughly 1.9 times their earlier rate since October 2025, according to an independent assessment published by AI safety nonprofit METR on November 5. The group said much of the acceleration appears to come from Anthropic’s growing use of AI in research and development, although it stressed that the estimate is uncertain and that the company redacted much of the evidence behind its conclusions.
The 11-week assessment ran from August 12 to October 29 and drew on confidential research, interviews and performance data from unreleased models. METR measured Anthropic’s benchmark progress using its version of the Epoch Capabilities Index and estimated that scores had grown at 1.9 times the rate seen before October 2025, with a 95% confidence interval of 1.5 to 2.8 times. The group said this would amount to roughly two years of historical benchmark progress over the past year, but warned that benchmark selection, model inclusion and other methodological choices affect the estimate.
METR guessed that a majority — perhaps about 80% — of the current acceleration as of October 2026 is attributable to AI, but emphasized that this judgment carries substantial uncertainty. Researchers reported that AI tools have helped with coding and experiments, but METR said the systems are not yet making highly significant, fully autonomous contributions to the most important stages of model training.
METR also said it found no clear evidence of near-term bottlenecks in algorithmic progress, training compute, experiment compute or data that would necessarily slow further development over the next 12 months. It acknowledged that its understanding of those constraints changed during the review and that it may have missed important limits. Anthropic was allowed to redact non-public evidence from the report, including some findings that METR said were important to its conclusions, limiting how fully outsiders can assess the results.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
