Anthropic is improving its models roughly twice as fast as it was a year ago, according to an analysis published Thursday by METR, the Berkeley-based nonprofit Model Evaluation and Threat Research, which Anthropic invited to review its internal data and interview staff.
The timing matters. The assessment landed days before Anthropic's planned Wall Street debut later this month — an offering that could be the largest IPO in history — and in the middle of an industry argument over whether AI development is accelerating beyond human control.
METR examined about eleven weeks of internal data, from mid-August through late October, plus employee interviews. Its conclusion, hedged with caveats, is stark: "Anthropic appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025." The group also reviewed both the public and a private version of Anthropic's August risk report and said the public one "overall understates the level of acceleration that was occurring."
The likely cause is the uncomfortable one. METR said it could confidently rule out growth in training compute — historically the main driver of capability gains — as the explanation. Its researchers' best guess is AI helping to build AI, the recursive self-improvement loop that Anthropic says its Claude models increasingly contribute to. The report found no bottlenecks in compute or data holding progress back, and said Anthropic staff described "signs of life" that AI could soon make significant, fully autonomous contributions to core parts of model training — while noting that is not the case yet.
METR was careful about the limits. The doubling may apply mainly to "well-specified tasks" that can be graded automatically rather than messy, hard-to-define work; Anthropic may also have understated earlier progress, flattering the current number. Redactions in the company's data limited how precisely METR could state its estimates, and its own team disagreed on some points.
The political context is sharper than the methodology. In a September 12 essay, CEO Dario Amodei called for the industry to "slow the pace at which we improve the capabilities of AI models," warning that recursive self-improvement "left unchecked, it could outrun our ability to understand and control these systems." He has also championed embedding outside evaluators inside labs and named METR as a candidate. This report is what that arrangement looks like in practice: a company-commissioned audit that finds the acceleration its own CEO warned about.
