METR published an independent assessment on November 5, 2026 concluding that, as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at about twice the rate seen between March 2024 and October 2025. Using Anthropic’s AECI (an Epoch Capabilities Index variant), METR estimates benchmark growth at 1.9\u00d7 the pre-October 2025 rate (95% CI 1.5\u20132.8\u00d7). The team attributes a majority of the current acceleration to AI-assisted R&D, rules out faster training-compute growth as the cause, and warns that missing bottlenecks raise the risk of further speedups \u2014 while stressing high uncertainty and heavy Anthropic redactions in the public report.<\/p>\n\n\n\n
METR (@METR_Evals)<\/a> announced the public report summarizing months of work inside Anthropic. The full write-up lives at metr.org<\/a>. Anthropic co-founder Sam McCandlish<\/a> replied that METR had sensitive internal data and independence, that many conclusions match his own, and that Anthropic redacted heavily because sharing “exactly what is driving the speedup” could hand rivals a roadmap. AP, The Verge and others carried the ~2\u00d7 framing with METR’s caveats attached.<\/p>\n\n\n\n Anthropic lists AI R&D acceleration among key risk indicators in its Responsible Scaling Policy. Outside observers usually see only system cards and public benches. METR’s engagement was scoped to whether AI is accelerating AI development, with access to internal benchmarks, training-domain compute metrics, staff interviews and related analyses \u2014 short of full employee-like roaming access or sampling non-public models. A compute-allocation verification piece was canceled mid-engagement; METR discloses that change.<\/p>\n\n\n\n METR fits trends on AECI scores for public and internal models (including unreleased “Model 2\u20134” snapshots dated by internal availability). A learned breakpoint around early October 2025 separates slower and faster regimes. Robustness checks across link functions, snapshots and benchmark subsets mostly keep acceleration near ~2\u00d7 (with some specs as low as 1.1\u00d7 or as high as ~2.9\u00d7). Interviews and Slack-poll uplift self-reports are treated as supportive but loose evidence. Fully autonomous contributions to the most important training steps are not yet dominant, METR says, though “signs of life” appear.<\/p>\n\n\n\nBackground: why this assessment exists<\/h2>\n\n\n\n
How METR measured acceleration<\/h2>\n\n\n\n