Anthropic’s AI Models Are Improving Nearly Twice as Fast, Independent Study Finds

A new study by METR, an independent AI safety research organisation, has found that Anthropic’s AI models are improving nearly twice as fast as they did previously. The research, published on November 5, examined the company’s progress using confidential research materials, interviews and performance data from unreleased models. Researchers estimated that the rate of improvement in Anthropic’s benchmark capabilities has accelerated significantly since October 2025.

According to the findings, Anthropic’s models have been improving at approximately 1.9 times their previous rate on well-defined tasks. However, METR stressed that the estimate carries considerable uncertainty, with the actual acceleration potentially ranging between 1.5 and 2.5 times the historical rate. The organisation also noted that the acceleration was not primarily driven by faster growth in computing power used to train the models.

The study highlighted the growing role of artificial intelligence in developing more advanced AI systems. Anthropic researchers have reportedly experienced significant productivity improvements by using AI tools to write code, conduct experiments and support research activities. Although some informal employee estimates suggested productivity gains of up to 12 times, METR cautioned that these figures were unreliable and should not be treated as established evidence.

The findings have renewed concerns about recursive self-improvement, a scenario in which AI systems increasingly contribute to building more capable versions of themselves. Researchers warn that rapid improvements could make advanced AI systems more difficult to monitor, control and align with human objectives. However, the study emphasised that AI has not yet demonstrated the ability to make highly significant contributions to the most important stages of model training entirely on its own.