Anthropic is now improving its AI models at about twice the pace it managed before late 2025, and most of that acceleration appears to come from the company using AI to help build AI. That is the central finding of an independent assessment published on November 5 by METR, a nonprofit that evaluates frontier AI systems, after its staff spent eleven weeks examining Anthropic's internal research process.
METR's team worked at Anthropic from August 12 to October 29, with access that the group described as unprecedented. The evaluators reviewed data on pretraining, reinforcement learning, training compute, benchmark scores for internal models, forecasts for a future model, and analyses of how much AI tools speed up research work. They also interviewed Anthropic staff. METR took no payment for the assessment.
The headline measure is AECI, Anthropic's version of the Epoch Capabilities Index, which combines benchmark results into a single score. METR estimates that since October 2025 those scores have grown at 1.9 times their earlier rate, with a 95 percent confidence interval of 1.5 to 2.8 times. In practical terms, the group says Anthropic made roughly two years of historical progress in the past year alone, about the distance from Claude 2.1 to Sonnet 4.5. The trend rose from about 15 points a year before October 2025 to about 29 points a year after it.
METR is careful to stress how uncertain that number is. It says it would not be surprised if the true acceleration were closer to 1.3 times or 3 times the historical rate, and it notes that the measure covers well-specified tasks, where progress may not translate evenly to broader, messier work.
One explanation METR could rule out was computing power. The group says it is confident the acceleration was not caused by faster growth in training compute. Its best guess is that a majority of the current acceleration, perhaps around 80 percent, is due to AI itself. Researchers told METR that AI is helping mainly as an assistant rather than through fully autonomous contributions, although there are signs that autonomous work could play a larger role soon.
METR also found no clear bottleneck that would prevent further progress over the next twelve months, whether in algorithms, training compute, experiment compute, or data. It warns, however, that its understanding of those bottlenecks changed several times during the review, so the finding carries real uncertainty. Even so, the absence of obvious limits raises the possibility that progress could accelerate further as better models help produce the next ones.
There is an important transparency caveat. Anthropic had the right to redact non-public information from the report, and METR says the company redacted a majority of the evidence it found most informative, describing the cuts as removing confidential information that could help rival developers. Readers therefore have to rely on METR's judgment more than the group would like. In a companion review, METR also said it is plausible that Anthropic had crossed its own early-warning threshold for dramatic AI-driven acceleration by mid-July, a claim Anthropic's August risk report had argued against.
For the wider industry, the report matters because it is a rare attempt to measure AI-assisted research from the inside rather than speculate about it. If the estimate is broadly right, the pace of model improvement is no longer set only by bigger computers and larger teams. It is increasingly set by the models themselves, and that feedback loop is now visible in the numbers.