Between August and October 2026, a small team from METR spent 11 weeks examining whether AI was accelerating research and development at Anthropic. They reviewed internal performance data for released and unreleased models, training compute, staffing data, and employee surveys. They also interviewed roughly 15 employees. Their assessment estimates that Anthropic is improving its models on well-defined tasks at roughly twice its previous pace. The researchers are confident that faster growth in training compute does not explain the acceleration. Their tentative estimate is that AI assistance accounts for about 80% of the current acceleration, although other factors likely played a larger role in late 2025.
There are important limitations. METR did not have unrestricted access to Anthropic’s internal systems or its unreleased models, and Anthropic redacted much of the evidence the researchers considered most informative. The estimated acceleration ranges from 1.1x to 2.9x depending on how the benchmarks are analyzed, and falls to 1.4x when models from other labs are included. Much of the evidence involves tasks with clear answers, so it’s unclear how well the findings apply to more complex real-world work.
Despite these qualifications, I think the central finding is significant. AI appears to be helping researchers build better AI faster, and METR found no obvious bottlenecks that would prevent further acceleration over the next year. This also brings two questions I raised recently about regulating frontier AI labs into sharper focus: what counts as AI-driven self-improvement, and what would meaningful independent evaluation require?