Anthropic Says AI May Soon Speed Its Own Research; METR Weighs In

Anthropic has warned that AI could soon dramatically accelerate the company's own research. Three days later, the evaluation group METR published an audit finding that Anthropic's pace of model development has roughly doubled.

The concept is known as recursive self-improvement, in which AI helps build more capable generations of AI. Anthropic's warning appeared in a revised IPO registration, while METR's audit was released on November 5, 2026, and covered by TechCrunch, The Decoder, and the Associated Press.

What the IPO Filing Says

According to TechCrunch, Anthropic wrote that its internal estimates and expectations for future models "indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI.'"

Wccftech describes that threshold as AI either replacing the entire research workforce at competitive cost or helping double the rate of progress in AI capabilities. Anthropic said: "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data."

The filing also admits that such acceleration may only be identified after it has happened. Anthropic said it does not anticipate being able to safely use Claude to fully autonomously design and build successive generations, and that it is considering policy updates to keep humans in control of frontier model development.

METR's Findings in Numbers

METR is a Berkeley-based nonprofit. The Decoder reported that it had access to Anthropic from August 12 to October 29, 2026, and took no payment from the company.

Figures are summarized from The Decoder's coverage of the METR report.

The Decoder also reported that METR considers it possible the warning level was reached as early as July 15, 2026. That goes further than Anthropic's own language in the filing.

The Caveats

METR cautioned that its results could be overstated, since the gains may apply mainly to well-specified tasks. Uncertainty is high, and METR said the public has to rely on its judgment more than would be ideal.

One reason is that Anthropic had redaction rights and used them to remove much of the evidence, including forecasts for future models. Anthropic said the redactions were aimed at protecting confidential information that could accelerate other AI developers. The Decoder added that Anthropic canceled the compute-allocation part of the audit midway through.

Views inside Anthropic also differ. One researcher put AI's share of productivity gains at just 15 to 20 percent. TechCrunch noted that an S-1 functions as both disclosure and marketing. Wccftech reported that Claude authored more than 90 percent of contributions to Anthropic's codebase in September, but that figure measures code volume rather than productivity.

What It Means for the Industry

The acceleration does not mean AI is building itself without people. Wccftech stressed that crossing the threshold would not necessarily mean achieving fully autonomous recursive self-improvement.

Even so, METR sees a significant risk that the acceleration continues over the next 12 months. It shares Anthropic's assessment that acute catastrophic risk is currently low, but it cannot rule out another sharp acceleration within six months. The central question now is whether safety work can keep pace, as the Associated Press reports Anthropic is planning to go public later this month.

Conclusion

Anthropic's warning and METR's audit suggest AI has begun to affect the speed of AI research itself. How much remains disputed, and some of the most important evidence is not public. The acceleration claims are best read alongside those limits.