METR Finds Anthropic’s AI Development Pace Has Roughly Doubled, Despite Redacted Evidence

Anthropic’s AI development advanced at roughly twice the historical rate during an 11-week evaluation by the Model Evaluation and Threat Research group, with AI use accounting for much of the acceleration. METR conducted the high-access assessment between August and October 2026, reviewing internal data and interviewing Anthropic staff.

The finding places Anthropic at the center of a larger question: whether AI systems are beginning to speed up the creation of more capable AI systems. The assessment focused on well-specified capability tasks rather than every aspect of frontier research. Coverage from The Verge, AP-linked reporting, and TLDR described the result as substantial while emphasizing that the conclusion carries significant uncertainty.

METR received unusually broad access to Anthropic’s research and development work, allowing its researchers to examine internal evidence rather than relying solely on public statements. The evaluation measured the pace at which Anthropic improved model capabilities on tasks with relatively clear definitions and benchmarks.

The review also included conversations with employees and analysis of company data, according to reporting published alongside the assessment. That setup gives the results more grounding than an outside estimate based only on model releases, while leaving the evaluation dependent on the scope of access Anthropic granted. Hacker News reproduced the assessment’s disclosure that Anthropic could redact non-public information, and The Verge described the investigation as separate from an earlier employee-like-access program for safety researchers.

Anthropic redacted a majority of the evidence METR considered most informative for its conclusions. The company said the removals were aimed at preventing confidential information from helping other AI developers reproduce or accelerate similar work.

That arrangement creates a tension at the heart of the assessment. METR was able to publish high-level conclusions about the pace of progress and the role of AI, but readers cannot inspect much of the underlying evidence. TLDR highlighted skepticism about the redaction-heavy presentation, while Hacker News focused on the difficulty of describing some withheld findings without exposing sensitive details. The limitation doesn't erase the result, but it makes independent scrutiny and replication harder.

The concern comes from the feedback loop: Anthropic uses AI systems to support research that produces stronger AI systems, which then assist with later research. METR’s estimate indicates that this loop is already associated with a marked increase in the pace of capability improvement at one major developer.

The result also arrives as Anthropic CEO Dario Amodei has called for caution about the speed of AI progress. AP-linked coverage framed the assessment against debates over human control and catastrophic risk, while The Verge reported that the company’s models were improving about twice as quickly as a year earlier with little indication of a slowdown. Those claims do not establish that AI development will continue accelerating indefinitely, but they raise the stakes for measuring the trend accurately.

The assessment shows a measured acceleration in Anthropic’s progress on selected, well-specified tasks, with AI contributing a majority of the observed gain. It does not establish that every part of AI research is advancing twice as quickly, or that frontier models will improve at the same rate across all domains.

The distinction matters because R&D speed depends on more than model capability. Compute availability, data, research priorities, evaluation design, engineering staffing, and organizational decisions all affect the output. Discussion summarized by TLDR also raised questions about real-world frontier-model regressions and broader existential risks that received less attention than the headline estimate. The redactions leave open how much the result depends on particular tasks or internal practices.

The immediate priority is better measurement. Independent evaluators will need comparable assessments across companies, time periods, and research tasks, alongside clearer disclosure about which evidence can be published without exposing competitive secrets.

Anthropic’s findings also intensify pressure on AI developers to document how models are used in model training, evaluation, coding, experiment design, and research planning. The assessment arrives as Anthropic prepares for a Wall Street debut, giving investors and policymakers a concrete data point in debates about growth, safety, and control. Its main significance is practical: AI-assisted R&D is now being treated as a measurable driver of frontier progress, not merely a productivity claim.