Anthropic AI Speed Doubles Despite CEO Cautions
Independent auditors found the company is advancing its models twice as fast as last year, contradicting its chief executive’s recent calls to slow down.
By Samuel Hayes ·
Dario Amodei told the world to slow down. The data says the opposite is happening.
In an essay published on Sept. 12, the CEO of Anthropic urged the industry to reduce the pace of model improvements so safety measures could catch up. He warned that unchecked recursive self-improvement could allow systems to outrun human control. Yet an independent analysis released Thursday reveals the company is moving faster than ever.
Model Evaluation and Threat Research, or METR, a nonprofit based in Berkeley, California, found that Anthropic is improving its AI capabilities roughly twice as quickly as it was a year ago. The assessment relied on internal data and staff interviews conducted over an 11 week period from mid August to late October. This acceleration is occurring just as the San Francisco based firm prepares for a stock market debut later this month, a listing that could become the largest initial public offering in history.
The speedup is driven largely by the use of AI to build better AI. METR researchers ruled out increased computing power as the primary cause. Instead, they attributed the gains to the company’s own models assisting in the development of newer versions. Amodei described this process as the growing ability of AI to build the next generation of AI. He noted in his September writing that this recursive loop must be pursued carefully.
Sam McCandlish, an Anthropic co founder and chief architect, responded to the findings in a post on X. He stated that many of the group’s conclusions align with his own views on the pace of research. McCandlish acknowledged that the company redacted parts of the report to protect commercial secrets. He explained that sharing exactly what drives the speedup risks handing competitors a roadmap to accelerate their own work. The company plans to release its own internal analysis of progress within weeks.
METR noted that Anthropic’s public risk report from August understated the level of acceleration. The independent group said it was given unprecedented access to internal data, which allowed them to identify specific trends. They found no significant bottlenecks in computing power or data holding back the progress. The report suggests that while AI is not yet making fully autonomous contributions to the core training process, researchers are seeing clear signs that this capability is emerging.
The findings carry weight because the Trump administration has encouraged AI companies to police their own safety standards. For years, details about model risks have come primarily from the companies themselves. METR is funded by donations and does not accept compensation for its evaluations, though it does receive tokens and access from the labs it partners with. The organization has evaluated models for Anthropic, OpenAI, and other major labs.
OpenAI CEO Sam Altman supported the concept of independent evaluators with employee level access at the same time Amodei championed the idea. METR’s report includes caveats, noting that the measured acceleration may apply only to well specified tasks that can be automatically graded. Despite these limitations, the group concluded that the rate of improvement has doubled, a finding that complicates the argument for a slower pace in the industry.
