METR estimated in a Nov. 5 assessment1 that Anthropic's benchmark progress since October 2025 runs about 1.9x its earlier rate, and said a majority of that acceleration is likely due to AI.
METR said that if AI systems start to significantly accelerate further AI development, this could lead to a sudden emergence of extreme and risky capabilities. It also said it did not identify clear bottlenecks that would stop the acceleration from compounding. Both findings are given with stated uncertainty, and Anthropic redacted much of the evidence behind them.
Background and access
METR published an October 2026 article, independent of this assessment, that proposed high-level questions about AI R&D acceleration it says the public needs answered to judge loss-of-control risks from frontier AI development. The Anthropic assessment applies internal evidence to those questions, according to the report1.
METR staff performed what the report calls a high-access evaluation at Anthropic from Aug. 12 to Oct. 29, 2026. METR said Anthropic shared internal data on pretraining, reinforcement learning, training compute by domain, benchmark scores for internal models, AI speedup analyses, forecasts of a future model's metrics and staff interviews. The Verge reported2 that METR said Anthropic shared more information than initially scoped, but that METR's access was narrower than the "employee-like" access Anthropic has promised other safety researchers in a separate program. METR said a small number of its data requests were denied.
The report is a METR blog post, not a peer-reviewed paper.
What METR found
The 1.9x figure comes from AECI, Anthropic's version of the Epoch Capabilities Index. METR said AECI scores for Anthropic's models have grown at 1.9x the pre-October 2025 rate (95% confidence interval: 1.5–2.8x), implying roughly two years of historical progress over the last year. Its headline judgment is narrower than the benchmark number: Anthropic is "likely" improving model capabilities on well-specified tasks at around twice the rate it did between March 2024 and October 2025. METR said it would not be surprised if the true figure were closer to 1.3x or 3x, and that it has additional uncertainty about how the acceleration translates to broader, less well-specified tasks.
Sensitivity checks across benchmark inclusion, model dating, model inclusion and benchmark saturation produced estimates between 1.1x and 2.9x, according to METR.
On cause, METR said it was confident the acceleration was not due to faster growth in training compute. It guessed that "a majority (perhaps ≈80%)" of the October 2026 acceleration is due to AI, a judgment it described as given with substantial uncertainty. METR said Anthropic's own staff did not reach a consensus on other causes, and that late-2025 acceleration was likely influenced by non-AI factors.
METR also said Anthropic told it that R&D is accelerated primarily through AI assistance rather than fully autonomous contributions, though there are signs the latter "will play a larger role soon." Staff self-reports of large productivity gains are consistent with the estimate, METR said, but are hard to interpret.
Limits METR states
METR said "most METR staff who worked on this project" hold the twice-the-rate view, and that the team still has significant uncertainty and internal disagreements. It said the AECI result "could be explained by overfitting to the benchmark distribution or otherwise be an artifact of the AECI methodology." It also said AECI data points lag research acceleration and are somewhat sparse, so the benchmark may not reliably capture recent changes.
METR listed factors that could push its estimate either way. It could be underestimating acceleration if it gave too little weight to further automation of AI R&D tasks, or to evidence that AECI has understated real-world impact. It could be overestimating it if it gave too little weight to non-AI factors, to AECI's focus on verifiable tasks, or to reasons Anthropic's 2023–2025 growth was unusually low.
On bottlenecks, METR said it saw no evidence of significant limits from algorithms, training compute, experiment compute or data over the next 12 months. It added that its understanding changed several times as it received more detailed information, so it is not confident in that assessment.
Redactions and Anthropic's response
Anthropic had rights to redact any non-public information in the post. METR said it was free to share its highest-level conclusions, but that Anthropic redacted a majority of the evidence it found most informative for them. Automatica reported earlier3 on that disclosure.
METR said much of the redacted evidence had multiple plausible interpretations, which makes it hard for the public to judge it independently. "As a consequence, the public has to rely on METR's judgment much more than would be ideal," the report said. METR said Anthropic described the redactions as "aimed at removing confidential information that could accelerate other AI developers." According to The Verge's reading of the report, METR said most of the redacted information was very high-level and that it was not in a position to say whether the redactions were justified commercially.
The Verge reported that Anthropic cofounder Sam McCandlish wrote on X: "We redacted a lot, and I accept that this makes it harder for others to check METR's work." He said the company determined that sharing what is driving the speedup "isn't the right thing to do," because it is commercially sensitive and risks handing others a roadmap to go faster. He said Anthropic plans to publish its own report on the speed of AI progress "in the coming weeks." According to The Verge, Anthropic declined to answer its questions and pointed to the post.