METR estimates Anthropic's model progress has roughly doubled in pace, mostly through AI use

METR staff (2.5 full-time equivalents) spent eleven weeks with access to internal Anthropic data. Most of them now believe that as of October 2026 Anthropic is improving model capabilities on well-specified tasks at "around twice the rate" of March 2024 to October 2025. METR says it would not be surprised by anything from 1.3x to 3x. Its one hard benchmark figure is that scores on AECI, Anthropic's version of the Epoch Capabilities Index, have grown at 1.9x the pre-October 2025 rate (95% confidence interval 1.5 to 2.8x). METR's own robustness checks move that between 1.1x and 2.9x depending on the link function used.

On attribution, METR guesses that perhaps 80% of the current acceleration is due to AI, "with substantial uncertainty". It says it is confident the acceleration did not come from faster growth in training compute. Anthropic staff self-reported uplift from AI reached 12x in Slack polls in October 2026, which METR calls only very loose evidence. It was told research is accelerated mainly through AI assistance rather than fully autonomous contributions, though researchers described early signs that autonomous contributions could grow. METR found no convincing bottlenecks over the next 12 months and says its view of the bottlenecks changed several times.

The limits are stated in the report itself. Anthropic redacted what METR describes as a majority of the evidence it found most informative, so METR says readers have to rely on its judgment. Ajeya Cotra, who writes Planned Obsolescence and says she did not work on the report, reads it as implying Anthropic's own acceleration threshold has already been crossed and extrapolates to superintelligence within a year. That extrapolation is hers. METR's report also lists reasons the true acceleration could be lower, including non-AI factors and a benchmark focus on verifiable tasks.

Why it matters: These are benchmark results on well-specified tasks, and METR says it is unsure how they carry over to messier work. Read it as a signal that model capability may keep moving faster than the 2024 and 2025 pace, which argues for building harnesses and verification that do not assume a model's current limits will hold for long. It does not measure what your own agents will do.