METR, a nonprofit that studies advanced AI, has reviewed how Anthropic's August 2026 Risk Report handles AI automating AI research. METR agrees with Anthropic that catastrophic risks from its models as of July 15 were low, but writes: "we think it is plausible that they had crossed their early warning threshold at the time." It also says the public report "understates the level of acceleration that was occurring."

Two terms, once

The early warning threshold is a tripwire in Anthropic's Responsible Scaling Policy. Anthropic's own report calls it "a potential early warning, rather than evidence that the threat has already materialized." Crossing it is a signal to look harder, not proof of danger.

2x acceleration is one of the policy's two Automated R&D triggers. It means AI progress running at double speed: as much progress in one year as would normally take two, in METR's reading of the policy. The other trigger is full substitution, a model that can replace Anthropic's entire research staff within a factor of five of the cost.

Where METR disagrees with the public report

METR found the public report's case on the 2x trigger "overstated." It also says most relevant METR and Anthropic researchers lean toward the line not having been crossed, though METR thinks it plausible it was. METR's main objections:

  • Leading indicators. The public report says acceleration was "by less than a factor of 2." METR calculated from the private report's own figures that the rate comes out at 1.97 (METR's math, not a number printed in that report) and says "around 2" is the better description. METR's own range was 1.3 to 2.8; Anthropic redacted that analysis.
  • How much is due to AI. The private report says progress is "slightly less than twice as fast" as it would be without AI, with "enough uncertainty that it seems plausible we do exceed this factor of 2." METR says the public report conveyed more confidence than that.
  • A one-time jump versus continued speedup. On Anthropic's capability index (AECI), the public report called a one-time jump "weak evidence." METR says the data cannot confidently separate a jump from continued acceleration. It puts the rate at about 1.5x with the data in the report and about 1.9x with benchmarks Anthropic added later.

Where METR agrees

On full substitution, METR finds Anthropic's argument "broadly compelling": the company tries hard to replace researchers with models, and researchers are still needed. Its confidence is lower than before, because key benchmarks have saturated, meaning they no longer measure further gains. METR also does not think the July 15 models could fully automate AI research.

The six-month warning

METR expects AI's help with research to keep growing. It writes that it "cannot rule out a substantial further acceleration in the pace of capabilities progress over the next 6 months," which "could rapidly lead to 'high' risks." In its separate assessment, METR says Anthropic is likely improving its models at around twice the rate of March 2024 to October 2025, with high uncertainty (it would not be surprised by 1.3x or 3x).

How independent was the review?

METR's published terms and disclosures say:

  • No money changed hands in either direction, and METR kept sole discretion over methods and conclusions.
  • Anthropic could redact non-public information and did. METR says it described every redaction that mattered to its conclusions, and that Anthropic redacted a majority of the evidence METR found most informative for its risk conclusions.
  • One METR team member had discussed a role with Anthropic, received an offer and declined it, and held equity in a competitor. Their work was checked by an unconflicted colleague.
  • Midway through, Anthropic dropped a planned check of how it splits compute between capabilities and safety. METR says that was before it reached any conclusions on it.

Anthropic shared confidential data from Aug. 12, METR delivered a draft Sept. 18, Anthropic approved the redacted report Oct. 29, and METR published Nov. 5.

Why it matters

The gap is the story. Anthropic's public report and its private one, as METR reads them, differ in how confident they sound about the 2x line, and the public reader sees the confident version. METR also says the redactions leave the public relying on its judgment "much more than would be ideal."

This lands as Anthropic's IPO paperwork turns to the same subject. Its amended filing says it "may imminently meet" the policy's accelerated-progress definition, while designating no current model over the line. In our reading, the filing talks about future models and METR about July 15, so they are not the same claim, and we have not checked that they use one definition. Nothing here speaks to the offering; our earlier look at the IPO filing numbers covers that paperwork.

What we checked

We read both METR posts in full and checked the quoted public-report wording against Anthropic's redacted Risk Report PDF. METR's full assessment holds the process disclosures. We did not see the private report or the redacted evidence, so every statement about them is METR's, and we found no Anthropic response in these documents. We link The Verge's coverage for that one point only.