METR Says Anthropic's Models Are Improving Twice as Fast, With AI Doing Much of the Research Work
METR, a nonprofit that studies advanced AI agents, has concluded that Anthropic's AI models are improving roughly twice as quickly as they were a year ago. The report, Independent Assessment of AI R&D Acceleration at Anthropic, was published on November 5, 2026.
METR describes the work as a high-access assessment of how far AI is accelerating AI research and development (R&D) at Anthropic. Its pilot evaluation involved 2.5 full-time equivalents of METR staff. Since August, a METR team had worked with internal Anthropic information over 11 weeks to examine whether AI itself is speeding up AI development.
On capability, METR estimated that Anthropic is improving model performance on well-specified tasks, meaning tasks with clear, checkable success criteria, at about twice the rate of a year earlier. METR also said benchmark scores, standardized tests run on identical problems, suggest Anthropic made roughly two years of historical progress over the last year.
METR said its researchers are confident the acceleration was not simply the result of scaling up compute, the raw processing power used to train models. It found that the majority of the acceleration as of October was due to AI, and that AI's contribution grew over time.
The mechanism matters as much as the magnitude. According to METR, much of the acceleration comes from AI assisting human researchers rather than working independently. METR also warned that there are signs fully autonomous AI contributions will play a larger role soon.
On productivity, Anthropic employees using the most capable internal model self-reported being, on average, around 12 times as productive as without AI assistance. METR gathered those figures through Slack polls and interviews. The 12x figure is around double what Anthropic staff reported a few months earlier while using Mythos 5. Reuters reported on the study's findings that Anthropic researchers use AI to write code and conduct experiments, with sharp productivity gains.
The report sits alongside earlier METR work on Anthropic's models. On September 22, 2026, METR published a predeployment evaluation, meaning a test run before public release, of Claude Opus 5.5. It assessed whether AI R&D at Anthropic would be dramatically accelerated by using that model. On the same day as the acceleration report, METR also published a review of the "Risks from automated R&D" section in Anthropic's August 2026 risk report. In that review, METR stated it agrees with Anthropic that by July 15th, internally deployed models posed low risk of directly automating R&D in key domains.
Anthropic was permitted to redact portions of the final METR report as part of its agreement with METR.
The benchmark-based estimates are measured outputs that other evaluators can, in principle, reproduce. The 12x productivity figure rests on self-reported poll responses and interviews, which capture how engineers perceive their own output rather than a controlled measurement of it. Both point in the same direction, but they are not interchangeable.
The broader context here is the distinction METR draws between AI assisting researchers and AI working autonomously. Assistance keeps humans in the loop on research direction and judgment. A shift toward fully autonomous contributions, which METR flagged as a possibility, would change how labs staff and supervise that work, and how external evaluators verify it. For practitioners, the report is a concrete data point on how fast an AI developer's own tooling can compound, and on the kind of high-access evaluation that makes such claims checkable at all.
