METR Estimates Anthropic's R&D Pace Reached 1.9x, with Most Acceleration Attributed to AI Assistance

On November 5, 2026, METR released a report stating that Anthropic's internal model capability growth rate (the AECI indicator) has accelerated since October 2025 to 1.9 times the historical rate, with a 95% confidence interval of 1.5 to 2.8 times.

Factual Reconstruction

The report covers an evaluation METR conducted at Anthropic between August 12 and October 29, 2026. Anthropic provided multiple categories of internal data, including pretraining, reinforcement learning, training compute, internal model benchmark scores, and AI acceleration analysis. The evaluation concluded that the acceleration did not come from a faster increase in training compute scale. Employee interviews and surveys showed that AI use was consistent with the doubling in the pace of capability progress. Anthropic requested redactions to most of the most informative evidence; METR can describe the redactions in the report but cannot disclose specific details.

Mechanism Breakdown

The report notes that the acceleration was mainly achieved through AI assistance to R&D workflows, rather than through fully autonomous contributions. Anthropic employees self-reported large productivity gains, but much of the detailed breakdown data was requested to remain confidential. The evaluation found no evidence that algorithmic progress, training compute, experimental compute, or data would form a significant bottleneck in the next 12 months, although METR said its confidence in the bottleneck judgment was limited because more detailed information had repeatedly changed its understanding. Some evidence allows multiple interpretations, and METR and Anthropic researchers are uncertain about how to synthesize it into an overall picture.

Industry Impact

If AI has become the main driver of frontier labs' acceleration of their own evolution, existing capability forecast timelines and safety evaluation cycles both need recalibration. The report emphasizes that the absence of obvious bottlenecks or strong diminishing returns means further capability progress could trigger even faster acceleration, but important limiting factors may also have been missed. The public currently relies more on METR's judgments than on direct access to the evidence, which reduces transparency.

Strategic Judgment

[Analysis] Based on the high-level conclusions METR has disclosed and the scope of confidentiality, the extent of AI-assisted R&D inside the lab may be faster than external observers can see. This would compress the time window for external evaluators to update baselines and require more frequent adjustments to safety evaluation designs. If similar evaluations are replicated at other labs in the future, the industry's overall speed of response to acceleration signals could become a key variable.