The METR AI acceleration assessment finds that Anthropic’s models have improved 1.9 times faster since October 2025 than during the preceding period. METR, a nonprofit AI evaluation group, reached that conclusion after spending 11 weeks inside Anthropic from August 12 to October 29, 2026. Its central measure was AECI, Anthropic’s in-house version of Epoch’s capabilities index. The result matters because faster model development could feed a loop in which improved AI helps researchers create better AI still, raising the risk of abrupt gains in dangerous capabilities.
METR AI acceleration assessment: A clear break in Anthropic’s progress rate
METR said AECI scores have risen 1.9 times faster since October 2025 than between March 2024 and October 2025, with a 95 percent confidence interval of 1.5 to 2.8 times. Before October 2025, Anthropic’s models gained about 15 AECI points per year. Since then, the rate has reached 29 AECI points per year.
The practical meaning is that Anthropic achieved in one year what previously took two. METR described the change as roughly comparable to the jump from Claude 2.1 to Sonnet 4.5. Most METR staff involved in the project now believe Anthropic was improving its models on well-specified tasks at around twice the rate recorded between March 2024 and October 2025 as of October 2026.
The headline figure carries meaningful uncertainty. METR said it would not be surprised if the true acceleration factor were as low as 1.3 or as high as 3. Robustness tests produced estimates between 1.1 and 2.9 times, depending on the method used. When models from OpenAI, Google and xAI were included, the industry-wide estimate fell to 1.4 times. METR said this could reflect Anthropic trailing competitors in 2025, with a catch-up effect distorting the comparison.
AI assistance appears to be the main driver
METR estimates that about 80 percent of the speedup comes from AI being used in Anthropic’s own research. That estimate comes with substantial uncertainty, but it points toward AI-assisted human work rather than fully autonomous research as the immediate explanation. METR found no evidence that AI was yet making highly significant and fully autonomous contributions to the most important parts of model training, although researchers described “signs of life” that such contributions could arrive soon.
Informal Slack polls from October 2026 showed employees saying on average that the latest internal model made them roughly twelve times more productive. METR called this very loose evidence. The polls were informal and may overstate gains. One Anthropic researcher instead estimated that AI accounted for only 15 to 20 percent of productivity gains and expressed distrust of colleagues’ self-reports.
METR was confident that the acceleration was not caused by a faster increase in training compute. It also found no bottlenecks in algorithmic progress, training compute, experiment compute or data that would slow the next 12 months. The organization qualified that finding because its understanding of the constraints changed several times as it received more information. Even so, METR believes there is a significant risk that acceleration will increase over the coming year.
Anthropic may be nearing its own warning level
Anthropic filed an amended S-1 warning on November 3 that internal projections indicate it may imminently meet its Responsible Scaling Policy threshold for dramatically accelerated AI-driven progress. That threshold requires progress to roughly double and for the doubling to be substantially attributable to automation of AI research. The 1.9 times AECI estimate alone therefore does not establish a policy breach.
The public August 2026 risk report said the acceleration was less than a factor of 2. The internal unredacted version implies 1.97, while Anthropic itself calls it plausible that the threshold has been crossed. METR also considers it possible that the warning level was already reached on July 15, 2026. However, METR noted that most relevant researchers at both organizations leaned toward the threshold not having been crossed at that point.
METR currently shares Anthropic’s assessment that acute catastrophic risk is low. It cannot rule out another sharp acceleration over the next six months, however, which could rapidly lift the risk rating. The distinction is important. A low assessment of immediate catastrophe does not mean that development controls are unnecessary, especially if the observed pace continues to compound.
Redactions limit independent scrutiny
Anthropic had unilateral redaction rights over all non-public information and used those rights heavily. Most of the evidence METR considered most informative was removed from the public version. The hidden material included forecasts for future models, breakdowns by training steps and the bottleneck analysis. Detailed acceleration estimates for stages of the training pipeline were also withheld.
One METR report image carries the words FAKE DATA in large letters where redacted information would otherwise appear. METR said this leaves the public reliant on its judgment much more than would be ideal and means the report does not achieve full transparency about the underlying evidence. Anthropic said the redactions were aimed at removing confidential information that could accelerate other AI developers.
The audit was not fully independent of practical constraints. METR took no payment from Anthropic, although Anthropic supplied laptops and token credits. Anthropic could approve or reject information requests, and most data arrived through two Anthropic technical staff, supplemented by about 15 interviews. METR also did not sample from non-public models. The full report runs to around 100 pages, and its researchers can publish risk findings without an Anthropic veto, but redactions still limit what readers can verify.
A canceled compute verification weakens the safeguards debate
Anthropic canceled the compute allocation audit midway through the project. The original scope included checking how Anthropic divides compute between capabilities work and safety work. Anthropic said it wanted the engagement focused on risk assessment and would revisit compute verification later.
Daniel Kokotajlo, the researcher behind the AI 2027 scenario, argued that auditing compute allocation is the foundation for making commitments to slow down verifiable. That argument matters because much of the public policy debate rests on the assumption that companies can measure and control the resources assigned to frontier development. A canceled audit leaves outside observers with less ability to check those commitments.
METR’s conclusion should therefore be read with two dimensions in view. The public data show a substantial measured acceleration in Anthropic’s benchmark performance, while METR also saw redacted internal evidence that it considered important. The central policy uncertainty is not whether a number exists, but whether outsiders can inspect enough evidence to determine how much of the acceleration comes from AI, whether it could compound, and whether Anthropic’s warning threshold should already apply.
For Claude users, investors and policymakers, the practical message from the METR AI acceleration assessment is that Anthropic’s development pace may have changed materially while remaining difficult to audit. The 1.9 times estimate, the roughly 80 percent AI contribution and the absence of obvious 12-month bottlenecks point toward a fast-moving research cycle. Heavy redactions, uncertain methods and the canceled compute verification mean those conclusions still require a meaningful degree of trust in METR’s judgment.