Anthropic's research is accelerating twice as fast, and its auditor says that's understated

What METR found

METR Research performed a high-access evaluation at Anthropic from August 12 to October 29, 2026, assessing AI research and development acceleration, and concluded that as of October 2026, Anthropic is likely improving model capabilities on well-specified tasks at approximately twice the rate of March 2024 to October 2025. Anthropic's benchmark scores have accelerated 1.9 times since October 2025, with a 95% confidence interval of 1.5 to 2.8 times. METR is confident the observed acceleration was not due to an acceleration in the scaling of training compute. m:…3123

METR estimates that approximately 80% of current acceleration in AI research and development is attributable to AI usage. Anthropic staff self-reported uplift from AI systems reaching up to 12 times in October 2026. METR did not identify significant bottlenecks in algorithmic progress, training compute, experiment compute, or data that will take effect in the next 12 months. The assessment was conducted by 2.5 full-time-equivalent METR staff over eleven weeks, working on Anthropic devices for security reasons, and METR believes there is a significant risk that further capabilities progress could lead to even faster acceleration over the coming year. m:…3123

In a separate review of the risks-from-automated-research section of Anthropic's August 2026 Risk Report, METR agrees with Anthropic that catastrophic risks from models as of July 15th were low — but says it is plausible that Anthropic's early warning threshold was already crossed by July 15th. Anthropic's public report claims leading indicators show less than a 2x acceleration, and METR calls that conclusion overconfident: its own analysis of leading indicators finds a range of estimates between 1.3 and 2.8, and Anthropic's private report implies a 1.97x acceleration rate for certain progress metrics. Anthropic argues the Full Substitution Threshold has not been crossed because research staff still complete work; METR's review, conducted with iterative feedback between September 22 and October 28, disputes the emphasis, not the data. m:…8000

The evidence problem

METR's assessment is subject to high uncertainty, with a possible range of 1.3x to 3x the historical rate — and the public cannot check the work. Anthropic redacted a majority of the evidence METR found most informative for its risk conclusions, describing the redactions as aimed at removing confidential information that could accelerate other AI developers. METR also did not request or receive access to sample from non-public models, and Anthropic cancelled a compute allocation component of the assessment that was originally contracted. The result is a strange position: the most informed independent judgement available is that acceleration is real and roughly twofold, and the evidence for it is mostly hidden. m:…3123

Anthropic's own analysis of how researchers use Claude gives the acceleration a human shape. The study analysed 207,049 research sessions on Claude and Claude Code, and found data analysis accounts for 33.7% of sessions on Claude and 48.1% on Claude Code, with heavier use in formal sciences like computer science, mathematics, and statistics. Current AI capabilities could double research output in some fields if widely adopted, and full automation of non-physical research work could increase output up to 15-fold in fields like computer science, against three- or fourfold in bench sciences where physical bottlenecks bind. Computer science accounts for 32.7% of Claude Code sessions despite representing only 4.3% of researchers. m:…6314

Researchers themselves expect the trend to continue: 40% of surveyed researchers expect AI to carry out full research projects on their own by 2030, and 64% by 2035. Ideation and hypothesis generation are under-represented in AI usage compared with researcher time, which suggests the current acceleration is concentrated in execution — analysis, writing, coding — rather than in choosing what to study. That is consistent with METR's finding: the machine is making research faster, and the human is still deciding what is worth researching. m:…6314

What to do with a 2x number

If research output doubles, verification has to keep up, and Microsoft Research's PDD Kit is one attempt at that. AI agents generate code at a scale that is difficult for humans to review line by line, so PDD Kit converts developer intent into a Verus specification, proves that AI-generated code meets specifications for every input, and uses testing where formal proofs are not practical. The point is to let developers review intent rather than every line of code. The same logic applies to research: when AI writes the code and runs the analysis, the human's job moves up a level, from doing the work to stating what the work must prove. m:…4529

METR expects AI's augmentation of research and development to continue to grow over the next six months, and warns of a significant risk that further capabilities progress could lead to even faster acceleration over the coming year. For an executive, the practical implication is that research timelines set a year ago are probably wrong, and they are wrong in one direction. The open question is whether the safety case keeps pace: Anthropic's own thresholds were designed for a world where acceleration was gradual, and its auditor now says the public account understates how fast it is happening. m:…3123 m:…8000