Anthropic may now be making two years of AI progress in one

From August 12 to October 29, the independent evaluation group METR examined Anthropic’s internal benchmark data, training-compute trends and research process, interviewed staff and worked with unreleased models. It estimates that Anthropic’s capability scores have risen 1.9 times as quickly since October 2025 as before, with a 95% confidence interval of 1.5 to 2.8. In the group’s interpretation, the company made about as much benchmark progress during the past year as it had during the preceding two.

METR is careful about what this measures. The index emphasizes well-specified, verifiable tasks rather than every messy activity involved in model research. Robustness checks produced acceleration estimates ranging from 1.1 to 2.9. Anthropic’s earlier progress may also have been unusually slow, which makes the change point look larger. This is therefore evidence of acceleration inside one laboratory, not a universal clock for the whole industry.

The more consequential attribution is that faster training-compute growth did not explain the change. METR’s subjective estimate is that a majority—perhaps around 80%—of the acceleration visible in October came from researchers using AI to do research. Staff described assistance rather than fully autonomous model development, but evaluators saw early signs that autonomous contributions could become more important. They found no clear algorithmic, training-compute, experiment-compute or data bottleneck likely to stop further progress over the next 12 months, while stressing that their understanding of those constraints changed repeatedly during the review.

The review also exposes the limits of laboratory transparency. Anthropic gave METR unprecedented access, but then redacted most of the evidence the evaluators considered most informative, arguing that disclosure could provide competitors with a roadmap for accelerating their own work. METR could publish its conclusions but not enough underlying material for outsiders to reproduce them. A separate review concluded that Anthropic’s August public risk report understated the acceleration then occurring and that its claim not to have crossed an early-warning threshold depended on contestable judgments.

That tension matters because CEO Dario Amodei has publicly called for the industry to slow capability development while Anthropic prepares for an IPO. The finding does not show a self-improving system operating without people. It does show a weaker but already compounding mechanism: better models make researchers faster, those researchers build better models, and the next research cycle begins from a higher level. The practical governance question is no longer whether AI-assisted AI research exists. It is whether independent evaluators can measure its pace early enough, with enough unredacted evidence, for pacing commitments to have operational meaning.