METR, an independent group that evaluates AI systems, concluded in a report dated 5 November 2026 that Anthropic is now improving its models about twice as fast as it did before late 2025. METR’s best guess is that most of the speedup comes from Anthropic’s own heavy use of AI in its research. The group spent eleven weeks inside the company, from 12 August to 29 October, with 2.5 full-time-equivalent staff on the job.
The headline figure rests on benchmark scores. Using AECI, Anthropic’s adaptation of the Epoch Capabilities Index (a single score built from many tests), METR estimates that scores have climbed 1.9 times faster since October 2025 than they did before, with a 95 percent confidence interval of 1.5 to 2.8 times. In practice, Anthropic made as much progress on those tests over the past year as it did across the two years before, a stretch METR places between the releases of Claude 2.1 and Sonnet 4.5.
Most METR staff on the project now put the current pace at about twice the historical one for well-specified tasks, and say they would not be surprised by 1.3 times or 3 times. They flag extra doubt over whether the gain carries over to messier, real-world work. Swapping the statistical assumptions behind the index moves the estimate anywhere from 1.1 to 2.9 times, and adding other labs’ models to the fit pulls it down to about 1.4 times.
On the cause, METR says it is confident the acceleration did not come from faster growth in training compute. After that, it found no decisive evidence, and Anthropic’s own staff did not agree among themselves. METR’s guess, offered with what it calls substantial uncertainty, is that perhaps 80 percent of today’s acceleration is due to AI, while the late-2025 jump was probably shaped by other factors.
The supporting evidence is soft. About 15 staff were interviewed, and METR had Claude sort their remarks: 2 clearly backed a doubling or more, 3 leaned that way, 4 were unclear, and 2 said no. One interviewee guessed AI explained only 15 to 20 percent and distrusted self-reported productivity. Slack polls of Anthropic staff showed self-reported speedups reaching about 12 times in October, which METR calls very loose evidence.
METR also found no convincing bottleneck in algorithms, data, or the compute used for training and for experiments over the next 12 months. It adds that its view of each bottleneck changed several times as more detail arrived, so it holds that finding loosely. Researchers described early signs that AI could soon make fully autonomous contributions to training, though today the gains come mainly from AI assisting people.
The access came with strings. Anthropic could approve or refuse any data request and held unilateral rights to redact anything non-public. METR says most of the evidence it found most informative was blacked out by Anthropic, which Anthropic said was meant to remove confidential information that could accelerate other developers. The public, METR concedes, has to lean on its judgment more than it should. Anthropic also dropped a planned check of how its compute is split between capabilities and safety partway through, saying it would revisit that later.
METR did not query Anthropic’s unreleased models, took no payment, and used laptops and free tokens supplied by Anthropic. It disclosed that one team member had been in talks about a job at Anthropic, received an offer, and declined; that person also held equity in a competitor, so a colleague re-checked their work.
The finding that matters for outsiders is how much of this sits out of view. METR’s figure draws on internal models that have not shipped, and the report says Anthropic holds recent models slightly longer before release. Anyone judging the pace from public releases alone is looking at a lagging picture.
Reported by METR in its assessment dated 5 November 2026.
