METR estimates Anthropic’s model development accelerated nearly twofold

METR conducted its high-access evaluation at Anthropic from August 12 to October 29, 2026.[1] The information it listed included pretraining, reinforcement-learning and training-compute metrics, benchmark scores, AI speedup analyses, forecasts and staff interviews.[1] The Associated Press reported that Anthropic invited METR to review internal data as the company prepared for a Wall Street debut.[2] Anthropic said it onboarded six METR employees with company hardware, office space and access to internal tools and systems.[3] This combination of internal access and multiple evidence types is important context for reading the result as an evaluation, rather than a single benchmark comparison.

METR concluded that Anthropic was likely improving models on well-specified tasks at about twice its March 2024-to-October 2025 rate, as of October 2026.[1] Its overall impression was a range of estimates centered around a twofold acceleration.[1] METR’s robustness analyses produced estimates from 1.1 to 2.9 times.[1] Its estimated 1.9-fold benchmark acceleration was higher than the 1.7-fold figure in Anthropic’s Haiku 5.5 system card, mainly because METR included newer unreleased models.[1] METR also said it would not be surprised if the current growth rate were closer to 1.3 or 3 times the historical rate.[1] These figures are estimates with meaningful variation, not a single precise rate.

METR estimated that about 80% of Anthropic’s current acceleration in October 2026 was due to AI, while emphasizing substantial uncertainty.[1] It said non-AI factors likely contributed to the acceleration Anthropic experienced in late 2025.[1] METR said the acceleration was not caused by faster growth in the scale of training compute.[1] It characterized Anthropic R&D as accelerated primarily by AI assistance rather than fully autonomous AI contributions.[1] Staff self-reports of AI-related uplift reached 12 times in October 2026, but METR described these as loose evidence because responses reflected low-effort consideration.[1] The reported figure should therefore not be treated as equivalent to METR’s broader estimate.

METR said Anthropic redacted a majority of the evidence it considered most informative for its risk conclusions.[1] It found no clear bottlenecks or steep diminishing returns, and said further capability progress could lead to even faster acceleration over the coming year, posing a significant risk.[1] Anthropic said METR could publish its analysis and that Anthropic could not veto publication, though it could require certain information to be redacted.[3] METR described AECI as Anthropic’s version of the Epoch Capabilities Index.[1] For readers assessing the findings, the redactions and the stated uncertainty are relevant limits on how confidently the estimates can be interpreted.