Artificial intelligence is increasingly helping researchers build more capable AI systems, and a new independent assessment suggests the process may be accelerating technological progress. Researchers at Model Evaluation and Threat Research (METR) found that Anthropic, the company behind the Claude AI models, is improving certain AI capabilities at roughly twice its previous pace. The findings, released November 5, offer a closer look at how AI tools are changing the development of the technology itself.
The assessment examined Anthropic’s internal research operations over approximately 11 weeks, including model performance data, employee interviews, and information about how AI is used during development. Researchers estimated that performance on a set of measurable tasks has been improving about 1.9 times faster since October 2025 than during the preceding period. They believe AI assistance accounts for much of that acceleration, although the exact contribution remains uncertain.
Rather than relying exclusively on human researchers to write code, analyze results, and conduct experiments, Anthropic increasingly uses AI systems to assist with those activities. This allows researchers to complete some development tasks more efficiently and potentially devote additional time to difficult problems. Importantly, the study did not find that AI systems are independently designing and training their own successors. Human researchers continue to direct the most consequential stages of development.
The findings also highlight questions about how quickly AI safety practices can adapt. If increasingly capable systems help researchers create better models in less time, the interval between major technological advances could shrink. METR cautioned that its estimates carry substantial uncertainty, particularly when applying results from narrowly defined tasks to more complex real-world capabilities. Anthropic also withheld some confidential information from the published assessment, limiting how much outside observers can independently evaluate.
The Readovia Lens
AI helping to develop AI represents a potentially significant shift in how the technology advances. Faster research could lead to improvements in medicine, science, software development, and other fields, but it could also make responsible oversight more challenging. The important distinction is that today’s findings describe AI-assisted research, not machines independently taking control of their own evolution. As these systems become more capable, understanding both their contributions and their limitations will be essential to evaluating what comes next.
