METR ran a high-access evaluation at Anthropic from 12 August to 29 October 2026 to assess how much AI is accelerating AI research and development. Its central finding is an approximately 1.9x acceleration in AECI benchmark scores since October 2025. METR says a majority of the current acceleration is likely attributable to AI use, though with significant uncertainty, and says Anthropic redacted a majority of the evidence it found most informative.
What METR found about the pace
METR estimates that Anthropic's AECI benchmark scores have grown at about 1.9x the pre-October 2025 rate, which it describes as roughly two years of historical progress in one year. The AECI trend went from +15 points a year before October 2025 to +29 points a year after, and the trend breakpoint was estimated at 7 October 2025. METR gives a 95% confidence interval of 1.5x to 2.8x, says it would not be surprised if the true acceleration was closer to 1.3x or 3x, and its robustness analyses show acceleration between 1.1x and 2.9x depending on methodology. Those are wide bounds around a single headline number.
The Haiku 5.5 system card reported 1.7x acceleration, and METR says its own higher figure comes primarily from including data on newer unreleased models. The internal models in the analysis are Unreleased 2022, Model 2 (July 2026), Model 3 (August–September 2026) and Model 4 (September–October 2026). When all frontier models are included, acceleration is lower, at 1.4x. Model 4 scores above 50% on all the AECI benchmarks it was tested on.
How much is down to AI
METR says a majority of the acceleration is likely attributable to Anthropic's use of AI to do R&D, and guesses, with substantial uncertainty, that perhaps 80% of the current, October 2026, acceleration is due to AI. It says it is confident the acceleration was not due to an acceleration in the scaling of training compute. In informal Slack polls, which METR considers only very loose evidence, Anthropic staff self-reported uplift from AI of approximately 12x in October 2026.
METR says AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process, but researchers described “signs of life” suggesting such contributions could come soon. It found no clear bottlenecks preventing further acceleration, and states there is a significant risk of even faster acceleration over the coming year.
What Anthropic redacted
Anthropic redacted a majority of the evidence METR found most informative. According to METR, Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers." METR also reports that Anthropic canceled the compute-allocation verification component of the assessment. METR says it did not take payment from Anthropic, though Anthropic provided free tokens for the assessment.
METR's Full Report is approximately 100 pages, and METR conducted approximately 15 interviews with key Anthropic staff. METR worked on Anthropic devices for security reasons, and it did not request or receive access to sample from non-public models. The assessment is distinct from the one Anthropic announced on 15 October. METR had published an earlier proposal on 14 October 2026 that framed the questions this assessment answers.
The limits METR flags
METR says its concern about underestimating acceleration includes the possibility of hyperbolic growth. Its concerns about overestimating include the possibility that non-AI factors explain most of the acceleration, that AECI focuses on well-specified tasks, and that Anthropic's early growth rate was unusually low.
Anthropic's filing
TechCrunch reported on 2 November that Anthropic's S-1/A, filed that day, states that its internal estimates and expectations regarding future models indicate it may imminently meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." It adds: "We have not designated any of our current models as exceeding this threshold, and we continue to assess the underlying data." The filing also states that Anthropic does not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, and a risk factor acknowledges it has in the past not identified, and may in the future not identify, a significant acceleration in its rate of development until after it has occurred. As Cotra quotes it, the Responsible Scaling Policy threshold is met when Anthropic observes or expects double the rate of progress compared with both the rate expected and the fastest rate of extended progress observed in the absence of significant AI contributions to AI R&D, and that it be plausible that the doubling is significantly attributable to automation of research and/or engineering.
Cotra's reading, and pushback
On 6 November, Ajeya Cotra, who calls the report's authors her METR colleagues but says she did not work on the project or see the unredacted report, wrote: "I think this report is the most important update we have gotten about AI risk in the last couple of years." She writes that, from her reading, it looks like the RSP threshold has already been crossed.
Cotra says her naive extrapolation implies superintelligence within a year: if acceleration is smooth, the rate of progress would double again within six months, then three months, then six weeks. Separately, she estimates, with great uncertainty, that researchers would need roughly a 10x labor uplift, which she says lines up with Anthropic staff self-reports of approximately 12x uplift. She also warns about weapons development and misalignment risks.
Cotra notes that in an earlier version of Anthropic's RSP, from May 2025, the highest named risk threshold for AI research automation corresponded to a somewhat weaker milestone, and that under the old RSP Anthropic would be required to have, at a minimum, security good enough to “protect against model-weight theft by state-level adversaries” by the time it reached that threshold. She writes that the AI industry is nowhere near this bar: “both OpenAI and Anthropic recently had major security breaches by amateur groups.” She expresses concern about transparency because of the redactions. On Hacker News, the commenter plaidfuji argued that it is a huge logical leap to go from accelerated research to AI systems that could develop novel weapons and overpower human civilization.
Sources
Bort, J. (2026) 'Anthropic warns that it's very close to RSI capability, aka self-building AI', TechCrunch, 2 November 2026, available at: https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/ (Accessed: 7 November 2026).
Cotra, A. (2026) 'AI is dramatically accelerating AI research right now', Planned Obsolescence, 6 November 2026, available at: https://www.planned-obsolescence.org/p/ai-is-dramatically-accelerating-ai (Accessed: 7 November 2026).
Hacker News (n.d.) 'AI is dramatically accelerating AI research', Hacker News, available at: https://news.ycombinator.com/item?id=50397994 (Accessed: 7 November 2026).
METR Blog (2026) 'How an assessor could investigate evidence of acceleration from the automation of AI R&D', METR Blog, 14 October 2026, available at: https://metr.org/blog/2026-10-14-ai-rd-acceleration-assessment-proposal/ (Accessed: 7 November 2026).
METR Blog (2026) 'Independent Assessment of AI R&D Acceleration at Anthropic in October 2026', METR Blog, 5 November 2026, available at: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/ (Accessed: 7 November 2026).
