AI · 7 November 2026 · 5 min read
Anthropic’s models sped up. Most of the evidence was cut
METR said on 5 November that, since October 2025, Anthropic’s benchmark scores have been rising at 1.9 times the earlier pace, with a range of 1.5 to 2.8. It guesses that about 80 per cent of the present speed-up is from AI used in the research. Anthropic cut a majority of the evidence METR found most useful. The public is being asked to trust the judgement.
METR published on 5 November an assessment of whether AI is speeding Anthropic’s own research. Staff were inside from 12 August to 29 October 2026. On benchmarks, using AECI, Anthropic’s version of the Epoch Capabilities Index, scores since October 2025 have grown at an estimated 1.9 times the earlier rate. The 95 per cent interval is 1.5 to 2.8 times. METR’s plain-language reading is that about two years of the old pace were packed into one year. A benchmark year is not a robot on a shift.
Separately, and with what it calls high uncertainty, METR says that as of October 2026 Anthropic is likely improving well-specified tasks at about twice the rate of March 2024 to October 2025. It would not be surprised if the true figure were 1.3 times or 3 times. Those are not the same number as 1.9. One is a fitted benchmark rate. The other is a judgement. METR guesses that a majority, perhaps about 80 per cent, of the October 2026 acceleration is due to AI. It is confident the cause was not a faster rise in the scale of training compute. It was told the speed-up is mainly assistance to people, not a lab running itself, and that AI is not yet making highly significant, fully autonomous contributions to the most important parts of training. For the most capable internal model in October, average self-reported uplift was about 12 times. Self-reported is the researcher’s own number.
What the public does not get to see
Anthropic redacted a majority of the evidence METR found most informative. The access deal already lets Anthropic require redactions. The cut categories include internal capability measures, algorithmic-progress estimates, part of the evidence on AI’s contribution, measures of research inputs, bottleneck and forecast detail, and part of the staff interviews. METR says it did not see evidence of a significant bottleneck, over the next 12 months, from algorithmic progress, training compute, experiment compute or data. That sentence is its reading of a file the public cannot check. No bottleneck found is not a promise that none exists.