Anthropic AI development speed doubles, study finds
An independent group says the company is improving its models twice as fast as a year ago, a finding that complicates calls to slow down the technology.
By Jonathan Reed ·
An independent research group says Anthropic is developing its artificial intelligence systems at twice the rate it was a year ago.
The finding comes from Model Evaluation and Threat Research, a nonprofit based in Berkeley, California. The group released its analysis on Thursday, November 5. The organization was invited by Anthropic to review internal data and interview staff. This review occurred as the San Francisco-based company prepares for an initial public offering later this month. The listing could become the largest stock offering in history.
The results stand out because Anthropic’s CEO, Dario Amodei, has recently urged the industry to slow down. In an essay published on September 12, Amodei called for a reduction in the pace of model improvements. He argued that safety measures need time to catch up with the technology. He specifically highlighted the risk of recursive self-improvement, where AI helps build the next generation of AI.
METR estimated that the acceleration in Anthropic’s capabilities is largely due to using AI to improve AI. The researchers analyzed data collected between mid-August and late October. They found that the company made as much progress on benchmarks in the past year as it did from October 2023 to October 2025. The group stated that it could confidently rule out increased computing power as the primary driver of this speedup. Instead, they attributed the gain to AI assistance in the development process.
Sam McCandlish, a co-founder and chief architect at Anthropic, responded to the report on the social media platform X. He noted that many of METR’s conclusions align with his own observations. He stated that the company plans to release its own analysis of AI progress within weeks. McCandlish explained that Anthropic redacted certain details from the report to protect commercially sensitive information. He argued that sharing exactly what drives the speedup could provide a roadmap for competitors.
The report includes caveats regarding the scope of the acceleration. METR noted that the faster progress may only apply to well-specified tasks that can be graded automatically. The group also acknowledged internal disagreements and uncertainty among its researchers. Despite these limitations, METR concluded that the rate of improvement had doubled. The organization said that Anthropic’s public risk report from August understated the level of acceleration occurring internally.
METR found no bottlenecks such as limits on data or computing power holding back the company. The researchers identified signs that AI systems are beginning to make autonomous contributions to the training process. Amodei’s essay described this trend as a key driver of recent advances. He warned that if left unchecked, this process could outrun human ability to control the systems.
