Anthropic Model Improvements Double, Study Finds

Anthropic is advancing its artificial intelligence models at roughly twice the speed it recorded a year ago. This acceleration occurs despite the company's CEO publicly urging the industry to slow its pace. The finding comes from an independent monitoring group that reviewed Anthropic's internal data just as the company prepares for a stock market debut.

The analysis was released Thursday by Model Evaluation and Threat Research, a nonprofit based in Berkeley, California. METR was invited by Anthropic to examine its internal records and interview staff. The company is scheduled to go public later this month, a move that could become the largest initial public offering in history.

The timing of the report creates a direct conflict with the stance taken by Dario Amodei, Anthropic's CEO. On Sept. 12, Amodei published an essay arguing that the AI sector needs to reduce the speed at which it improves model capabilities. He stated that the industry must slow down to allow safety measures to catch up with the technology's rapid evolution.

METR estimated that the increase in speed is largely driven by AI itself helping to build newer versions of the technology. The group analyzed data from mid-August through late October, covering a period of about 11 weeks. Researchers concluded that Anthropic made as much progress on benchmarks in the past year as it did between October 2023 and October 2025.

Sam McCandlish, a co-founder of Anthropic and its responsible scaling officer, responded to the report in a post on X. He wrote that many of METR's conclusions match his own views on the speed of research. McCandlish noted that while AI is speeding up the work, it is difficult to quantify exactly how much. He added that the company plans to release its own analysis of progress within the next few weeks.

The nonprofit stated that it could confidently rule out increased computing power as the cause for the acceleration. Traditionally, more compute leads to better model performance. METR researchers believed the improvement stems from using AI to assist in the development process. They noted that there are no current bottlenecks in computing power or data holding back the work.

McCandlish explained that the collaboration with METR began before the company committed to embedded third-party evaluators. He described the partnership as an illustration of why external evaluators are useful. The company redacted parts of the report to keep commercially sensitive information private. McCandlish said this tradeoff was necessary to avoid giving competitors a roadmap for achieving similar speedups.

The report included caveats about the data. METR noted that the measured acceleration might apply only to well-specified tasks that can be graded automatically. The group also found that Anthropic's public risk report understated the level of acceleration occurring internally. METR is funded by donations and does not charge for its work, though it receives access and tokens from the labs it evaluates.

Amodei's September essay described recursive self-improvement as the primary driver of recent AI advances. He defined this as the growing ability of AI to build the next generation of AI. The CEO warned that if left unchecked, this process could outrun human ability to understand and control the systems. OpenAI CEO Sam Altman supported the use of independent evaluators with similar access during that same period.

METR found that AI is not yet making fully autonomous contributions to the most critical parts of model training. However, Anthropic researchers described signs that this capability is approaching. The independent group emphasized that the rate of progress determines how quickly risks from advanced models increase. The company's upcoming public offering will place its financial and operational performance under greater scrutiny from investors and regulators.