METR Says Anthropic May Already Have Crossed Its AI Research Threshold

Anthropic's own filing says its next models may "imminently" meet the company's threshold for AI-driven research acceleration, and the outside reviewer METR said Thursday that the threshold may already have been crossed. The two documents, published three days apart, disagree about timing but not about direction.

Anthropic disclosed the language in an amended S-1, its IPO registration statement, filed Monday, TechCrunch reported. Julie Bort, TechCrunch's venture editor, quoted the filing as saying internal estimates "indicate that we may imminently meet" the Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." The policy is Anthropic's published rulebook for what it must do as its models get more capable.

What the Nov. 2 filing says and does not say

The filing is careful. "We have not designated any of our current models as exceeding this threshold," it says, and it adds that Anthropic "may in the future not identify" an acceleration until after it has happened.

Anthropic also said it does not expect to use Claude to design and build successor frontier models without meaningful human oversight. It is considering policy updates to "keep humans in control of frontier model development at Anthropic."

What METR found on Nov. 5

METR, a nonprofit that evaluates frontier models, published a review of the automated-research section of Anthropic's August 2026 Risk Report. Its conclusion: "we think it is plausible that they had crossed their early warning threshold at the time," meaning Anthropic's models of July 15. It added that "we believe the public report overall understates the level of acceleration that was occurring."

METR's own leading-indicator analysis gave "a range of reasonable estimates, between 1.3 and 2.8" times the pace without AI. Anthropic's private report, which METR reviewed with redactions, puts end-to-end progress at "slightly less than twice as fast." METR calculated that this implies about 1.97 and said "around 2" describes it better than "less than a factor of 2."

METR also said that "most relevant researchers at METR and Anthropic lean toward the threshold not having been crossed." That lean is the strongest evidence for Anthropic's reading, and METR put it in the same review.

1. METR, low end1.3 x
2. Cotra, as of October1.8 x
3. Anthropic private report, METR's calculation1.97 x
4. METR, high end2.8 x

Source: METR review, Nov. 5, 2026 (1.3, 1.97, 2.8); Ajeya Cotra, Planned Obsolescence, Nov. 6, 2026 (1.8). Different dates and methods; the threshold is 2x.

Cotra reads it as crossed

Ajeya Cotra, who writes the Planned Obsolescence newsletter, argued on Nov. 6 that the threshold "has already been crossed." Her best estimate is a 1.8x speedup relative to no AI as of October, which she thinks has grown since. She did not work on the METR project and did not see the unredacted report.

She also cited about 10x research-labor uplift as what the faster capability pace would require, against about 12x in Anthropic staff self-reports, and called the 10x figure "very uncertain." A naive extrapolation, she wrote, "implies that we would reach extreme superintelligence within a year." She flagged that as naive herself.

The sceptic's case

Not everyone reads autonomous AI research as that capable. Arjun Panickssery wrote in a comment on his LessWrong shortform that "I struggle to believe at an intuitive level that autonomous R&D is so ingenious." In his own projects the model located a relevant blog post but "didn't realize its importance," and he wrote that its judgment is "very poor in a basic way."

His exception is "certain hill-climbing tasks with fast feedback loops," where he wrote that repeated prompting improved a Scrabble move generator by 10x. That is a single anecdote, and METR's evidence is mostly redacted.

What to watch

Anthropic's filing is a registration statement and the METR review covers models from July 15, so neither says what the company's newest models do. METR said it "cannot rule out a substantial further acceleration in the pace of capabilities progress over the next 6 months," which it said could rapidly lead to "high" risks.

The same tools are already in developers' hands: see how Claude Code handles user data and what an audit of its sessions can and cannot catch.