Anthropic’s artificial intelligence models are improving nearly twice as fast as before, partly because AI is helping researchers develop more advanced systems, according to an independent study.

The findings were published on November 5 by AI safety nonprofit Model Evaluation and Threat Research (METR). The organisation examined how much AI is accelerating the development of future AI models.

The 11-week assessment ran from August 12 to October 29, 2026. It included confidential company research, staff interviews and performance data from unreleased models.

METR found that Anthropic’s model benchmark performance has improved at approximately 1.9 times its previous rate since October 2025.

The organisation said its estimate carries significant uncertainty. Its analysis placed the likely acceleration between 1.5 and 2.8 times the historical rate.

AI Is Playing a Growing Role in AI Development

METR believes AI tools are contributing significantly to Anthropic’s faster research progress.

Researchers found that AI is helping employees write code, conduct experiments and complete other research tasks more efficiently.

Anthropic employees also reported substantial productivity gains from using AI tools. Informal internal surveys indicated average self-reported productivity improvements of around 12 times for the company’s most capable internal model in October.

Microsoft Makes Windows 11 Faster on PCs With 8GB RAM

However, METR cautioned that these figures are not reliable measurements of overall research productivity. Self-reported estimates can be difficult to verify and may overstate actual gains.

The study also found that AI is not yet making major, fully autonomous contributions to the most important stages of model training.

Faster Progress Raises AI Safety Concerns

The findings have renewed concerns about recursive self-improvement, a scenario in which AI systems increasingly contribute to developing more capable versions of themselves.

Such a development could accelerate technological progress beyond the pace of traditional human-led research.

METR warned that faster AI development could create significant risks if safety assessments fail to keep pace. More capable systems could become harder to monitor, control and align with human objectives.

The organisation also found no convincing evidence of major bottlenecks in algorithmic progress, training compute, experiment compute or data that would significantly slow development over the next 12 months.

Importantly, METR said the acceleration was not caused by faster growth in the amount of computing power used for model training.

Anthropic’s Findings and the Limits of the Study

Anthropic provided METR with access to internal research data, performance measurements and interviews with employees.

However, the company exercised its right to redact much of the evidence considered most informative. METR acknowledged that these redactions make it harder for the public to independently assess some conclusions.

The study therefore carries important limitations, particularly around the precise contribution of AI to the company’s research progress.

Anthropic did not immediately respond to a request for comment, according to Reuters.

The findings add to a growing debate over whether AI companies can maintain effective safety oversight as their systems become increasingly involved in their own development.