METR said in a 5 November 2026 assessment that Anthropic used its redaction rights to remove a majority of the evidence METR found most informative for its risk conclusions about Anthropic's AI research process.
METR wrote that the public version "does not achieve 'transparency about evidence'"1, so outside readers must rely on METR's own judgment on a question it says bears on whether AI could drive a "sudden emergence of extreme and risky capabilities."
What METR says was redacted
In a redaction summary at the top of its post, METR wrote that1 Anthropic "had rights to redact any non-public information," while METR had "the right to disclose if Anthropic exercised redaction rights, and whether or not any of our findings depended significantly on redacted information."
METR then wrote that while it could "freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions." It added that "some of the key findings that most informed our conclusions about risk were redacted by Anthropic," including its more detailed estimates of AI's cumulative and instantaneous effects on Anthropic's rate of progress.
METR quoted Anthropic as describing the redactions as "aimed at removing confidential information that could accelerate other AI developers." Anthropic's position is documented only through that quotation.
METR said much of the redacted evidence "had multiple plausible interpretations," which makes it "hard for the public to judge this evidence themselves." It concluded that "the public has to rely on METR's judgment much more than would be ideal."
METR's post refers to a "Full Report" behind the public version. A post on X by Casey Newton2, quoting what it presents as METR's statement, describes it as a "~100-page full report" submitted to Anthropic. The post is a secondary account, and the figure does not appear in the METR text reviewed for this article.
Background
METR said staff performed a "high-access evaluation" at Anthropic from 12 August to 29 October 2026 to assess how far AI is accelerating AI research and development. The data it said it reviewed included pretraining and reinforcement learning metrics, training compute, benchmark scores for internal models, AI speedup analyses, forecasts and staff interviews.
In an earlier review of the misalignment section of Anthropic's Risk Report (August 2026)3, METR described its work as keeping the public informed about catastrophic AI risks, including from models used only inside companies. It said it conducts third-party reviews of company risk reports at their request.
That October review said Anthropic could veto publication or redact any part of it, but "no redactions were made" and METR did not change its conclusions. METR said it made minor changes after Anthropic feedback, limited to wording, clarification and improving the methodology of an experiment.
METR called that review "relatively lightweight" and based solely on the internal Risk Report. Its authors were engaged before an announcement titled "An update on independent evaluators" and "did not have the level of access to information described in that post." METR said it expects "future engagements to be governed by different terms."
Even there, METR recommended publishing four pieces of information Anthropic had redacted from its own Risk Report. It said some "substantially informed our views on risk pathways," and that examples of past concerning model behavior led it to see Anthropic's monitoring system as weaker than public evidence suggested.
Automatica reported on 6 November4 on a forum post summarizing the acceleration finding. That coverage said the underlying claims and data were unverified and did not address the redactions.
The acceleration findings
METR wrote that most staff on the project believe Anthropic is, as of October 2026, improving model capabilities on well-specified tasks at around twice its March 2024 to October 2025 rate. It cited "significant uncertainty and internal disagreements" and said the true figure could be closer to 1.3x or 3x.
Benchmark scores measured by AECI, Anthropic's version of the Epoch Capabilities Index, have grown at an estimated 1.9 times the earlier rate since October 2025, with a 95% confidence interval of 1.5 to 2.8x. Robustness checks gave 1.1x to 2.9x.
METR gave two separate caveats on AECI. The data points lag research acceleration and are sparse, and the result could reflect overfitting to the benchmark distribution or an artifact of the AECI methodology.
METR said it is confident the acceleration was not due to faster growth in training compute. It guessed that a majority, "perhaps ≈80%," of the current acceleration is due to AI, a judgment it described as made with "substantial uncertainty."
METR also said it saw no clear bottlenecks or steep diminishing returns in the next 12 months, but that its understanding changed several times and it "may be missing important bottlenecks."
METR said a table in its process section lists the evidence it was and was not able to discuss. Automatica did not review that section or the table, so the full scope of the redactions is not reflected here.