METR Says Anthropic May Have Crossed Its Own 2x AI Research Acceleration Warning Line

In a review published on November 5, METR agreed with Anthropic that the risk of catastrophe from automated AI research was low for the models Anthropic had as of July 15. It added: "we think it is plausible that they had crossed their early warning threshold at the time." The threshold in question is reached when AI doubles the speed of Anthropic's progress. METR was also critical of the company's public report: "we believe the public report overall understates the level of acceleration that was occurring."

This matters because the leading AI labs have said openly that they aim to automate AI research itself. METR's review is one of the few outside checks on how close any lab is to that goal, and the reviewers saw evidence that Anthropic has not published.

Public and Private Numbers

Anthropic's public report puts the acceleration at "less than a factor of 2." METR says the company's private, unredacted report implies a figure of about 1.97x. METR worked that number out itself; it does not appear in the private document. METR's own analysis produced a range of 1.3x to 2.8x, depending on how acceleration is defined and which period is measured. METR said the measurement uncertainty is large, and Anthropic redacted METR's discussion of it.

The two documents also differ on how much of the speed-up comes from AI. The public version, METR wrote, implies fairly high confidence that AI accounts for only a smaller share. The private version estimates that end-to-end progress is "slightly less than twice as fast" as it would be without AI, and says it is plausible the factor exceeds 2. METR said it broadly agrees with the private assessment and that the public report should have said so.

On AECI, Anthropic's internal index of model capability benchmarks, the estimated acceleration was about 1.5x on the original data. It rose to about 1.9x after Anthropic added more benchmarks, using the same set of models. An unreleased model, called Model 2 in the review, scored about 1.5 points above Claude Mythos 5, with large error bars. METR said the leading indicators fit sustained acceleration better than a one-off jump, but that it cannot yet tell the two apart and that future measurements should settle the question.

On a second threshold, the full substitution of research staff by AI, METR found Anthropic's case broadly convincing because there is no visible sign of researchers being automated away. It said it held that view with less confidence than before, because few reliable benchmarks remain unsaturated.

The review took more than two months to clear. Anthropic shared its unredacted report on August 14 and METR sent its first draft on September 18. Anthropic approved the redacted version on October 29. The redactions cover the specific progress rates, METR's own leading-indicator analysis and its discussion of measurement uncertainty. METR said every redaction that matters to its conclusions is described in the document.

A Call for a Ban

The review has already been taken up in Washington's debate over AI. In an X post on November 6, New York Assemblymember Alex Bores cited it while urging Congress to ban recursive self-improvement, meaning AI systems that build their successors, "this year if possible," according to an account of the post by FourWeekMBA. Bores wrote that OpenAI says it is on track to fully automate AI research by March 2028, that Anthropic employees cite 2027, and that "Any mistake or misalignment in current AIs could be multiplied in each generation." His post did not name a bill. His term in the Assembly ends on January 1, 2027.

METR said the catastrophic scenarios in Anthropic's threat model would probably need much more acceleration than existed on July 15. It also said it cannot rule out substantial further acceleration over the next six months, which could quickly push the risk to "high".