METR, an independent AI evaluation group, says Anthropic is likely improving its models’ abilities on well-specified tasks at about twice the rate it managed between March 2024 and October 2025, and that AI tools are probably behind most of the speed-up. METR staff worked inside the company with high access from 12 August to 29 October and published a public summary of a roughly 100-page report on 5 November.
On METR’s benchmark trend, growth since October 2025 has run at 1.9 times the earlier rate, and METR says anything from about 1.3 to 3 times would not surprise it. It is confident the change was not driven by faster growth in training compute and estimates that about 80% of the acceleration is due to AI, with substantial uncertainty. It found no convincing bottleneck to slow progress over the next 12 months, though with low confidence.
AI that speeds up its own development is one of the scenarios safety researchers and governments watch most closely, and this is a rare outside measurement of it. The limits matter too: Anthropic had redaction rights and removed much of the evidence METR found most informative, so readers must rely on METR’s judgement, and the gains may be confined to tasks that are easy to grade automatically. METR says no fees were paid in either direction.
Source: METR