METR independent assessment published November 5, 2026, after an 11-week review, August 12 to October 29, 2026
Twice the Pace, Mostly Redacted: METR Finds AI Speeding Up Anthropic's Own Research
METR, the nonprofit Model Evaluation and Threat Research group, published an independent assessment on November 5 concluding that Anthropic is likely improving its models' capabilities on well-specified tasks at about twice the rate it managed between March 2024 and October 2025. The estimate rests on 11 weeks of work inside Anthropic, from August 12 to October 29: access to pretraining, reinforcement learning, compute, and internal benchmark data, speed-up analyses, and about 15 staff interviews, with METR taking no payment. Measured on AECI, Anthropic's version of the Epoch Capabilities Index, benchmark scores have grown at an estimated 1.9 times the earlier rate since October 2025, with a 95 percent confidence interval of 1.5 to 2.8 times, which METR frames as roughly two years of historical progress in one year. METR says it is confident the acceleration was not caused by faster growth in training compute, and guesses that perhaps about 80 percent of it is due to AI itself, while stating plainly that the attribution carries substantial uncertainty. It found no evidence that AI can yet make highly significant, fully autonomous contributions to the most important stages of model training, and no clear bottleneck in algorithms, compute, or data likely to slow progress in the next 12 months. The limits on the record are part of the finding: Anthropic held redaction rights over non-public material and, in METR's words, redacted a majority of the evidence it found most informative; a planned check on how much compute goes to capabilities versus safety work was dropped from scope midway. In a companion review, METR says Anthropic's claim not to have crossed a 2x acceleration threshold rests on contestable judgment calls, that crossing it is plausible, and that Anthropic's public August 2026 Risk Report overall understates the acceleration that was occurring.
Executives
Read this as a planning input with its confidence attached. The strongest independent evidence yet says a frontier lab's development pace has roughly doubled in a year, mostly because the lab uses its own product to do the research, and the evaluator found no near-term bottleneck. Every roadmap in your organization that assumes today's AI capability levels hold for the next planning cycle should be stress-tested against a faster curve. At the same time, the most informative evidence behind the headline is redacted, so treat the precise figure as an estimate from a credible evaluator working with partial public proof, not as a measured constant.
CIO / CISO
The governance lesson is in the companion review: a developer's public claim that a self-defined acceleration threshold has not been crossed rested, in the outside reviewer's judgment, on contestable calls, and the public report understated the underlying pace. Apply that pattern to your own AI governance. Vendor safety and risk reports are inputs, not verdicts; where a deployment is material, require the evaluation method, the threshold definitions, and the evidence standard behind any not-crossed claim, and record who besides the vendor has reviewed it. METR's access model, months inside, primary data, staff interviews, no fee, is the reference point for what independent actually looks like.
Engineers
The technical content worth your time is the measurement design: a composite capability index tracked over time, a rate change estimated against a defined baseline period, and an explicit attempt to separate the compute explanation from the AI-assistance explanation before attributing the acceleration. Note also what the acceleration is made of so far: assistive use at scale, researchers testing many more ideas, rather than autonomous contributions to core training, per both METR and the staff interviews it published. That distinction is where realistic near-term expectations for your own teams should be set.
Everyone (non-IT)
Headlines about AI improving itself can sound like science fiction, so here is the measured version. An outside nonprofit spent 11 weeks inside one AI company and concluded its models are getting better about twice as fast as they used to, largely because its researchers use the AI heavily in their own work. The same report says the AI is not yet doing the most important parts of the job on its own, and that much of the proof could not be published. Both halves are the story, and the second half is why the first half deserves a careful rather than a panicked reading.