An independent safety group found that Anthropic's AI models have been improving about 1.9 times faster since October 2025 than in the preceding 18 months, and estimated that roughly 80 percent of that speedup comes from Anthropic using its own AI to do its research. The nonprofit Model Evaluation and Threat Research, known as METR, published the finding on Thursday after nearly 11 weeks of what it called unprecedented access to Anthropic's internal data, from August 12 to October 29. Anthropic held unilateral redaction rights over the public report and used them to strip out most of the evidence METR considered its most informative.
METR measured the acceleration with AECI, a composite index that bundles many benchmarks into a single capability score. Anthropic's scores rose about 15 points a year before October 2025 and about 29 points a year after, which METR read as a 1.9x acceleration, with a stated range from 1.3x to 3x. The group said it could confidently rule out faster growth in training compute as the cause, leaving the use of AI to build AI as its best explanation. One Anthropic employee told the auditors they were "now only bottlenecked on ideas and compute, no longer bottlenecked on code."
The number matters because Anthropic's own Responsible Scaling Policy treats a doubling of progress driven by automated AI research as a warning threshold. In its August risk report, covering the period through July 15, Anthropic said acceleration was "less than a factor of 2." METR called that case overstated: the internal, unredacted figures implied about 1.97x, and METR said it was plausible the threshold had already been crossed by mid-July, while noting most researchers at both organisations lean toward it not having been crossed. Anthropic's amended IPO filing, dated November 2, now says future models "may imminently meet" that same definition of dramatically accelerated progress.
The redactions are the report's central weakness, and METR said so plainly. Anthropic removed the detailed stage-by-stage acceleration estimates, the bottleneck analysis and the forecasts for future models, describing the cuts as "aimed at removing confidential information that could accelerate other AI developers." One chart in the public version carries the words "FAKE DATA" where real figures were pulled. METR wrote that the public must now rely on its judgment "much more than would be ideal." Anthropic also cancelled, midway through, a planned audit of how it splits computing power between capability and safety work.
The findings sharpen a tension AI Daily has tracked in recent issues. In September, Anthropic chief executive Dario Amodei urged the industry to "slow the pace" of capability gains and championed embedded external evaluators, naming METR specifically; last week more than 1,100 AI workers signed a declaration calling for limits on automated AI research. This audit is the first hard measurement of the thing those warnings describe, and it arrives weeks before Anthropic's planned public listing. METR cautioned heavily: the 1.9x figure swings between 1.1x and 2.9x depending on one modelling choice, falls to 1.4x when rival labs' models are included, and the benchmarks used to detect acceleration are themselves saturating, which could make the next jump harder to see. METR currently agrees with Anthropic that acute catastrophic risk is low, while warning it cannot rule out another sharp acceleration in the next six months.