Artificial intelligence is increasingly becoming part of the process used to develop more advanced artificial intelligence systems. A new independent assessment by AI evaluation nonprofit METR suggests that this shift may already be accelerating progress at Anthropic, the company behind the Claude family of large language models.
Published on November 5, 2026, the assessment estimates that Anthropic's AI model capabilities on well-defined tasks have been improving at approximately twice their previous rate. METR's analysis found that the growth rate of Anthropic's benchmark capability index increased by an estimated 1.9 times after October 2025 compared with the earlier period. The findings raise new questions about how AI-assisted research could influence the future development of large language models.
METR Finds a Faster Rate of AI Model Improvement
METR conducted an 11-week assessment at Anthropic between August 12 and October 29, 2026. The research involved access to internal development metrics, benchmark results from unreleased models, interviews with employees and information about the company's research and training processes.
According to the report, Anthropic's benchmark capability index, known as AECI, grew approximately 1.9 times faster after October 2025 than during the earlier comparison period. The statistical estimate has a 95% confidence interval of 1.5 to 2.8 times. METR's broader judgment is that the improvement rate on well-specified tasks is approximately twice its previous pace, although the organization emphasizes considerable uncertainty around that assessment.
The finding does not mean Anthropic's models suddenly became 1.9 times more intelligent. It describes the estimated rate of improvement over time on a collection of benchmark tasks, not a direct measurement of general intelligence or overall product performance.
AI Assistance May Be Accelerating AI Research
One of the most significant conclusions concerns how Anthropic uses its own AI systems during model development. Researchers found evidence that AI tools are helping employees with coding, experiments and other research activities. METR believes that AI assistance likely accounts for a majority of the recent acceleration, although other factors also contributed.
This creates a potentially important feedback loop. More capable models can assist researchers with parts of the development process, helping teams conduct experiments and improve software more efficiently. Those improvements can contribute to newer models, which may then provide more capable assistance to researchers.
However, METR distinguishes AI-assisted research from fully autonomous AI development. Its assessment does not establish that AI systems can independently design, train and deploy their own successors. Human researchers continue to play an important role in the most significant stages of model development.
Why the Findings Come With Important Limitations
The assessment includes uncertainty about both the scale and causes of the acceleration. METR says its broader estimate could plausibly be closer to 1.3 times or three times the historical rate. The organization also notes that much of the evidence it considered most informative was redacted by Anthropic because of confidentiality and commercial concerns.
These redactions make it harder for outside researchers and the public to independently examine some of the evidence behind the conclusions. Although METR was given substantial access to internal information, readers cannot independently reproduce every part of its assessment using publicly available data.
The researchers also said that the observed acceleration was not explained by a faster increase in training compute. This suggests that improvements in research methods and the use of AI tools may be important factors, rather than simply increasing the amount of computing power used to train models.
What This Could Mean for the Future of Large Language Models
If AI-assisted research continues to improve, developers could potentially shorten parts of the model development cycle and experiment with new techniques more quickly. This could affect competition between AI laboratories, the pace of large language model releases and the resources required to develop increasingly capable systems.
It also creates challenges for AI safety evaluation. If model capabilities advance more quickly, independent evaluators and developers may have less time to assess new systems, identify risks and improve safeguards before deployment.
For now, METR's assessment is evidence of a possible acceleration in AI-assisted research, not proof that recursive self-improvement has arrived. The distinction matters: AI systems helping people develop better models is a meaningful technical development, but it is not the same as AI independently creating increasingly powerful generations of itself.
What Happens Next?
The relationship between AI capabilities and AI-assisted research is likely to remain an important area of investigation. Further independent assessments, more transparent evidence and comparable measurements across different AI laboratories will be needed to understand whether the trend continues.
For the large language model industry, the central question is no longer only how quickly researchers can build more capable AI systems. It is also how much those systems can contribute to building the next generation, and whether safety evaluations can keep pace with that progress.
