Anthropic AI development speed doubles
A nonprofit found the San Francisco company is advancing its models twice as fast as a year ago, contradicting its CEO's call to slow down.
By Paula Delgado ·
An independent research group found that Anthropic is improving its artificial intelligence systems at roughly twice the rate from 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. It contradicts public statements by Anthropic CEO Dario Amodei, who has urged the industry to reduce the pace of technological development.
In September, Amodei wrote an essay responding to concerns about AI control. On Sept. 12, he told the sector to "slow the pace at which we improve the capabilities of AI models." He argued that safety measures need more time to keep up with progress. Amodei specifically recommended using external evaluators with deep access to company operations. He named METR as a suitable third party for this role.
METR researchers spent about 11 weeks reviewing Anthropic's internal data. This period ran from mid-August to late October. The team also interviewed Anthropic staff. The company granted the researchers access to internal data and provided tokens for their work. METR said it received "unprecedented access" to Anthropic's research and development processes. The group does not take compensation for its evaluations, relying instead on donations and access agreements.
The report states that Anthropic made as much progress on benchmarks in the past year as it did between October 2023 and October 2025. METR attributed this acceleration primarily to the use of AI to build better AI. The group ruled out increased computing power as the cause. Traditionally, more compute leads to better performance, but METR found no such correlation here. Anthropic's Claude chatbot is now helping to develop newer versions of the system.
METR noted that no bottlenecks, such as limits on data or compute, are currently holding back progress. This creates a risk of compounding acceleration. The report says AI is not yet making fully autonomous contributions to model training. However, Anthropic researchers described "signs of life" suggesting this capability is approaching. Amodei called this process recursive self-improvement. He warned that if left unchecked, AI could outrun human understanding and control.
Anthropic is preparing to go public later this month. The company is based in San Francisco. It may stage the largest initial public offering in history. Anthropic did not immediately respond to requests for comment on the new findings.
