METR estimates Anthropic's AI progress on well-specified tasks has roughly doubled

As of 9:10 p.m. ET, Monday, Nov. 9, 2026

Anthropic is likely improving its AI models on well-specified tasks at around twice the rate it did between March 2024 and October 2025, and most of that speedup probably comes from the company's own use of AI in research, the AI safety nonprofit METR said in an assessment published Thursday, Nov. 5.

METR stressed that the estimate carries "high uncertainty," saying it "would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."

What METR found

METR staff spent 11 weeks, from Aug. 12 to Oct. 29, reviewing internal Anthropic data, including benchmark scores for unreleased models, training-compute figures and interviews with staff, the report says. METR said it did not take payment from Anthropic.

Benchmark scores for Anthropic's models have grown about 1.9 times faster since October 2025 than before, METR said, with a 95% confidence range of 1.5 to 2.8 times. That implies "roughly two years of historical progress over the last year," the group wrote.

METR said it was "confident" the speedup was not caused by faster growth in the computing power used for training. It estimated that "perhaps ≈80%" of the current acceleration is due to AI, while saying that figure is highly uncertain. Staff told METR that AI mainly assists human researchers and cannot yet make highly significant, fully autonomous contributions to the most important parts of model training, though they described "signs of life."

The group said it found no clear bottlenecks likely to slow progress over the next 12 months, creating "a significant risk" of even faster acceleration, though it cautioned it may be missing some.

Caveats and redactions

Anthropic redacted "a majority of the evidence" METR found most informative, METR said, describing the redactions as aimed at protecting information that could help rival developers. As a result, the public "has to rely on METR's judgment much more than would be ideal," the report says.

METR also said the speedup might be limited to tasks that are easy to grade, and could be overstated if Anthropic's earlier progress was unusually slow. Anthropic dropped a planned review of how it divides computing power between capabilities and safety work midway through the project, METR said.

Anthropic's response

Anthropic co-founder Sam McCandlish said the company plans to share "a report with our own analysis of the speed of AI progress within Anthropic," The Verge reported. The Verge noted that CEO Dario Amodei published an essay in mid-September titled "We Must Pace the Frontier."