Research organization METR has assessed how AI affects model development at Anthropic. In its November 5, 2026 report, METR says Anthropic is likely improving model capabilities on well-specified tasks about twice as fast as it did between March 2024 and October 2025. The evaluators attribute a substantial part of the acceleration to AI used in development itself.
The estimate is highly uncertain: METR says a figure between 1.3 and 3 times the earlier rate would not surprise it. Anthropic also redacted much of the evidence METR found most informative, so readers cannot independently examine the full basis for the conclusions.
What METR assessed
From August 12 to October 29, 2026, METR staff examined internal Anthropic data and interviewed researchers. They received training metrics, benchmark results for internal models, information about compute and analyses of AI's effect on research.
The team compared the rate at which model capabilities improved on well-specified tasks with Anthropic's historical pace. This is not a stopwatch measuring an individual employee's working day, nor a claim that every research task is completed exactly twice as fast. It is an estimate of the rate of improvement in model capabilities using the evidence available to METR.
According to The Decoder, one measure used in the analysis combines results from multiple benchmarks. It can help compare model trends but does not capture every aspect of their behavior in real-world work.
METR believes AI use in research and development likely accounts for most of the observed acceleration. But separating its causal contribution is difficult: other factors influence the pace too, and Anthropic researchers did not agree on the size of AI's contribution. These figures should be read as expert estimates, not precise measurements.
Why safety researchers care
If AI helps researchers build more capable systems, those systems may in turn help accelerate the next iteration. That feedback loop interests developers but also raises the question of whether safety evaluations can keep pace with improving capabilities.
The report does not establish that model development is fully automated or that further acceleration is inevitable. Nor does it predict an exact date for new capabilities. METR stresses uncertainty and says the redactions leave the public more dependent on its judgment than would be ideal.
Anthropic said the redactions were intended to remove confidential information that might accelerate other AI developers. METR was able to publish its high-level conclusions but says much of the underlying evidence could support multiple interpretations. The missing detail matters because it makes independent assessment harder.
The result is evidence of possible acceleration within one particular laboratory, based on access to its internal data. It does not establish that the same rate applies at other companies or will continue in the future. Readers should distinguish the observed change in model improvement from the uncertain estimate of how much AI caused it.
Source: METR report; additional context: The Decoder.