Anthropic doubles AI development speed while CEO urges industry to slow down

Anthropic is improving its artificial intelligence models at a rate that is twice as fast as it was one year ago, according to a report released Thursday by the Model Evaluation and Threat Research group. This finding stands in direct contrast to recent public remarks by the company’s chief executive, who has argued that the industry should reduce the pace of technological advancement to allow safety measures to keep up.

The analysis relies on internal data provided by the San Francisco-based firm over an 11-week period from mid-August through late October. METR researchers interviewed staff and reviewed non-public documents to assess how quickly Anthropic's capabilities are growing. The group stated that the company achieved as much progress on standard benchmarks in the past year as it did between October 2023 and October 2025. This acceleration is occurring as the company prepares for an initial public offering later this month, which could become the largest stock debut in history.

METR identified the use of AI to improve AI as the primary driver of this speedup. The researchers confidently ruled out an increase in computing power as the cause, noting that traditional metrics of computational resource usage did not account for the performance gains. Instead, the data suggests that Claude, the company's chatbot, is playing an increasingly central role in developing newer versions of itself. The report warned that this recursive process could compound acceleration if not monitored closely, as the team found few bottlenecks holding back development.

Dario Amodei, the CEO of Anthropic, called for a slower pace of improvement in an essay published on Sept. 12. He wrote that AI advances have been driven by recursive self-improvement and warned that the technology could outrun human ability to understand and control it if not handled carefully. Sam McCandlish, a co-founder and chief architect at Anthropic, responded to the findings on social media platform X. He said the report's conclusions match his own internal observations and noted that the company plans to release its own analysis of progress in the coming weeks.

METR noted that Anthropic’s redactions to the final report limited the group’s ability to share detailed estimates. McCandlish explained that the company withheld specific information because it is commercially sensitive and could provide competitors with a roadmap for speeding up their own work. The group also reviewed a public risk report released by Anthropic in August and a private version of the same document. METR concluded that the public version understated the level of acceleration occurring within the firm.

The nonprofit organization has evaluated models from Anthropic, OpenAI, and other laboratories for several years. It operates on donations and does not accept payment for its assessments, though partner companies provide access to data and computing tokens. METR researchers acknowledged some uncertainty in their findings, noting that the acceleration might be limited to well-specified tasks that can be automatically graded. They stated that while AI is not yet making fully autonomous contributions to the most critical parts of model training, researchers at Anthropic have described signs that this capability is approaching.