METR Finds Anthropic Model Gains Running at Nearly Twice the Earlier Pace

The capabilities of Anthropic's models have advanced at about 1.9 times the earlier rate since October 2025, according to an independent evaluation by the nonprofit Model Evaluation and Threat Research group, which estimates that roughly 80 per cent of the acceleration may be linked to the company's use of AI in its own research and development. METR attaches a high degree of uncertainty to that figure. The finding bears directly on the contested question of recursive self-improvement — whether increasingly capable systems can contribute to building their successors — a scenario the report says has not yet been reached in fully autonomous form.

Over an 11-week assessment running from August to October 2026, a team of six METR researchers worked from Anthropic's San Francisco headquarters, interviewing employees, examining internal data and analysing the performance of models not released to the public. Staff described using AI tools to write code, conduct experiments and carry out development work. Internal polling put the average self-reported productivity gain at roughly 12 times for the most capable internal model as of October, a figure METR treats as weak evidence, since respondents may not represent the workforce and may have answered carelessly. One researcher told the group that work was "now only bottlenecked on ideas & compute, no longer bottlenecked on code". METR judged that additional computing power in training does not explain the acceleration, and stressed that most of the gain still comes from AI assisting people rather than performing the most important stages of training independently; it found no evidence of highly significant, fully independent contributions at those stages.

Anthropic finished training three unreleased models in the last six months — completed in July, around August or September, and around September or October — and METR separately lists an internal-only model from 2022. The report estimates the company covered "roughly two years of historical progress over the last year". Keeping state-of-the-art models private gives product and research teams a commercial advantage, though executives may also withhold them on safety grounds; METR's Ajeya Cotra told Business Insider earlier this year that she expected AI safety's "most serious incidents" to occur at Anthropic and OpenAI.

METR reported no clear bottlenecks in computing power, data availability or experiment design that would necessarily block further advances over the next 12 months, while cautioning that the assessment is uncertain and that important constraints may have been missed. The group wrote that the absence of observed bottlenecks or strong diminishing returns means there is a significant risk that further progress could lead to even faster acceleration, and said it had seen signs that fully autonomous contributions will play a larger role soon, without dating that shift. Evidence remains limited: METR said that while it could freely share its highest-level conclusions about risk, Anthropic redacted a majority of the evidence it found most informative for those conclusions.