Independent evaluator METR says Anthropic is improving models on well-defined tasks at about twice its 2024 to 2025 pace.
The big picture: METR spent August 12 through October 29 inside Anthropic's research process, reviewing internal model metrics, training data, compute trends and staff interviews. The group concluded that capability progress on well-specified tasks has accelerated sharply.
- Its central benchmark trend showed a 1.9x acceleration since October 2025, with a 95% confidence interval of 1.5x to 2.8x. METR said a true rate closer to 1.3x or 3x would still be plausible.
Between the lines: METR is confident that faster growth in training compute did not cause the change. Its researchers estimate, with substantial uncertainty, that AI tools may account for roughly 80% of the current acceleration.
- Anthropic staff described the gains as AI assistance to human researchers rather than fully autonomous model development. Evaluators also saw signs that autonomous contributions could grow soon.
Yes, but: Anthropic redacted much of the evidence that most influenced METR's risk conclusions. The benchmark emphasizes verifiable tasks, its data points are sparse, and METR said changes unrelated to AI could explain part of the result.
What's next: The evaluators found no clear compute, data or algorithmic bottleneck likely to stop further progress during the next 12 months, though they changed their assessment several times as Anthropic disclosed more information.
The bottom line: The report does not show an AI system independently building its successor. It does show a frontier lab using AI throughout the research pipeline while the measured pace of improvement rises.
Why it matters: If AI materially accelerates AI research, safety reviews, regulation and competitors all get less time to respond between generations of increasingly capable models.
