Anthropic’s Benchmark Scores Rise at 1.9 Times Their Previous Pace as AI Aids Research
METR estimates that Anthropic’s model benchmark scores have been rising faster. AI’s main role was to assist researchers, rather than conduct research independently, and its contribution has not been definitively quantified.
AI safety research organization METR released a report on November 5 (local time) assessing the pace of AI research and development at Anthropic. It estimated that Anthropic’s model benchmark scores have been improving about 1.9 times as fast since October 2025 as they did before. The analysis included scores from internal models that have not yet been released. METR concluded that a substantial share of the acceleration likely came from using AI in research and development.
What improved 1.9 times as fast?
The figure compares the rate at which Anthropic’s models improved their benchmark scores. From August 12 to October 29, METR reviewed internal model test scores, training metrics and staff interviews. The aim was to examine how using AI in research relates to improvements in the next generation of AI.
The estimate comes with a range. METR reported a 95% confidence interval of 1.5 to 2.8 times the previous rate. The benchmark scores used here primarily reflect performance on tasks that are easy to grade. Whether those gains translate into better performance on everyday tasks requires separate evaluation.
How does AI help researchers?
The changes METR observed mainly involved AI assisting researchers. METR concluded that AI independently carrying out core research from start to finish was not the main driver. The report’s central finding is that incorporating AI into human research workflows accelerated improvements in model performance.
METR also examined whether the acceleration could be explained by faster growth in the computing resources used to train models. It concluded that faster growth in training compute was not responsible for the acceleration. However, it did not definitively quantify AI’s contribution and noted substantial uncertainty about the influence of other factors.
Why has this sparked talk of “self-improving AI”?
The prospect of more capable AI helping develop the next generation of AI drew attention. The Verge explored why AI-assisted AI development raises concerns among researchers. AI Times focused on the possibility of “recursive self-improvement,” in which AI repeatedly helps create more capable AI. Commentary on GeekNews also suggested that repeating this process could lead to continued acceleration.
It is difficult to pin down AI’s contribution numerically. METR said a substantial share of the acceleration likely resulted from using AI in research. The Decoder reported that METR estimated AI use accounted for roughly 80% of the acceleration observed as of October 2026. METR noted substantial uncertainty around that estimate. It also acknowledged considerable uncertainty about how long the observed pace could continue.
What Korean users should watch
This is an evaluation report, not a product announcement. Anthropic removed from the public version most of the evidence that METR identified as most important to its conclusions. That makes it difficult for readers to examine the underlying evidence themselves. Separately, on October 15, Anthropic announced a collaboration giving METR access comparable to that of an employee.
The findings show that AI is being used not only to assist with work but also as a tool for researching the next generation of AI. For Korean users and businesses, the key question is whether the rapid gains seen in internal tests will be noticeable in Korean-language tasks and real-world work. Future evaluations will be worth watching to see whether this pace holds and whether the benefits appear in the products people use.