METR says Anthropic understated AI R&D acceleration risk in its August 2026 report
Article timeline
The articles in this story. Times are ET.
Andrew CurranXPickMETR estimates AI drives most of Anthropic's recent capability accelerationAIMETR published a risk assessment estimating that Anthropic has probably reached the 2x acceleration threshold in its Responsible Scaling Policy, and may be running as high as 3x. METR guesses about 80% of the acceleration as of October 2026 comes from AI itself. The report says its benchmark estimate has wide uncertainty, with a 95% confidence interval of 1.5 to 2.8x.
METROfficialMETR reviews Anthropic's automated R&D risk section in its August 2026 reportAIMETR publishes a review of the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report. The post links to the full review on METR's blog.
- METROfficialMETR publishes AI R&D acceleration risk assessment using internal Anthropic information
AIMETR says it has published a report on whether AI is accelerating AI development, based on a risk assessment that used internal Anthropic information since August. Anthropic could redact nonpublic information and METR kept the right to disclose when it did, and Anthropic made some redactions, which are marked in the review.
- METROfficialMETR says models as of mid-July posed low risk of automating AI R&D
AIMETR says even a roughly 2x acceleration would likely fall short of the catastrophic scenarios Anthropic describes. It agrees that the risk of models directly automating R&D was low as of mid-July.
- METROfficialMETR says Anthropic's public report understated its acceleration risk
AIMETR says the key question is whether Anthropic had crossed its acceleration threshold at the time, and that Anthropic's report concluded it had not. METR reviewers' best guess was also no, but METR says the public report understated how plausible it was that acceleration was above the threshold.
- METROfficialMETR reviews Anthropic's claims on automated AI R&D risks
AIMETR says it reviewed the "Risks from automated R&D" section of Anthropic's August 2026 Risk Report, using a nonpublic version. The review focuses on the report's claims about catastrophic risks from automated AI R&D and whether Anthropic had crossed relevant capability thresholds.
- METROfficialMETR describes access, redactions, and limits in its Anthropic report
AIMETR says its report covers the access it had, the redactions Anthropic made, and the limits on its work. The report also includes an AEF-1 disclosure, conflict-of-interest information, and a summary of its contract terms.
- METROfficialMETR says it may be underestimating AI acceleration
AIif it underweighted three signals: recent increases in some growth rates, early signs of fully autonomous contributions, and cases where AECI understated real-world impact.
- METROfficialMETR says it may be overestimating AI acceleration in three ways
AIMETR says it may be overestimating acceleration if it underweighted non-AI factors in Anthropic's research progress. It also flags AECI's focus on verifiable tasks and reasons Anthropic's unusually low early growth rate in 2024–2025.
- METROfficialMETR sees no convincing 12-month bottlenecks in AI progress, but may be missing some
AIMETR says it did not see convincing evidence of significant bottlenecks within the next 12 months in algorithmic progress, training compute, experiment compute, or data. The organization says its view changed during the assessment, so it may be missing some bottlenecks.
- METROfficialMETR says AI capability acceleration did not come from faster compute growth
AIMETR says its full analysis estimated how much of the acceleration in AI capabilities came from compute, data, and algorithmic progress. METR says it is confident the acceleration was not caused by an increase in the growth rate of training compute.
- METROfficialMETR finds Anthropic model benchmark scores growing 1.9x faster since October 2025
AIMETR reports that since October 2025, AECI scores for Anthropic's models have grown at 1.9x [95%: 1.5–2.8] the earlier rate. The post says analysis choices affect this estimate and that it remains unclear how benchmark scores map to real-world capabilities.
- METROfficialMETR says AI progress may accelerate over the next 12 months
AIMETR says it believes capabilities progress could accelerate faster in the next 12 months because it has not observed clear bottlenecks or strong diminishing returns. The organization adds that its limited vantage point means it may have missed factors that could slow progress.
- METROfficialMETR says Anthropic redacted most key evidence from its public assessment
AIMETR says it submitted a roughly 100-page full report on its assessment to Anthropic. Anthropic redacted a majority of the evidence METR found most informative from the public version, so the public must rely more on METR's judgment than METR considers ideal.
- METROfficialMETR finds Anthropic's nonpublic AI R&D evidence highly informative
AIMETR says Anthropic shared extensive nonpublic evidence across its AI R&D pipeline, from pretraining to RL, including estimates of algorithmic progress over time. METR calls this evidence highly informative about AI R&D acceleration.
- METROfficialMETR publishes assessment of whether AI is accelerating AI development
AIMETR says a team within the organization has worked since August with internal Anthropic information to assess whether AI itself is accelerating AI development. METR is now publishing a report summarizing the results of this risk assessment.
- METR BlogOfficialPickMETR says Anthropic understated AI R&D acceleration before its July 15 models
AIMETR reviewed the automated R&D section of Anthropic's August 2026 Risk Report. It agrees that July 15 models posed low catastrophic risk, but says they plausibly crossed the early-warning threshold for 2x acceleration. METR argues the public report understates the evidence for acceleration in leading indicators and AECI.