METR report: Anthropic's AI models are speeding up its own research
AI safety research organization METR published a report on November 5, 2026 saying Anthropic’s AI models are now improving roughly twice as fast as they were a year ago, and that most of that acceleration comes from AI itself rather than more compute. The finding, from an 11-week probe inside Anthropic, matters because it is some of the clearest outside evidence yet that a leading lab’s AI is already speeding up the lab’s own research.
What’s new
According to METR’s report:
- Anthropic’s benchmark scores suggest the company “has made roughly two years of historical progress over the last year,” and METR says it is confident the acceleration isn’t simply from scaling up training compute.
- Employees using Anthropic’s most capable internal model self-reported being about 12 times as productive at their work as they would be without AI help, roughly double what staff reported just a few months earlier using an internal model METR calls Mythos 5.
- Most of the acceleration so far comes from AI assisting human researchers rather than working independently, though METR says there are “some signs that fully autonomous contributions will play a larger role soon,” and it found no clear signs of diminishing returns.
- Over the past six months, Anthropic finished training three unreleased models that METR codenames Model 2, Model 3 and Model 4, completing training in July, August–September and September–October respectively, plus an older unreleased model from 2022.
- Anthropic had the right to redact parts of METR’s roughly 100-page full report before publication, and METR says it “redacted a majority of the evidence we found most informative.” Anthropic also canceled part of the assessment that would have evaluated how much compute it puts toward safety work versus capabilities.
Anthropic cofounder Sam McCandlish responded in a post on X, as quoted by The Verge, writing: “We redacted a lot, and I accept that this makes it harder for others to check METR’s work,” but that sharing exactly what is driving the speedup “isn’t the right thing to do,” citing commercial sensitivity and the risk of “handing out a roadmap for others to go faster.” He said Anthropic plans to publish its own report on the pace of AI progress “in the coming weeks.”
What it means for developers
The report doesn’t give developers a new API or tool to try, but it’s a data point on where frontier labs are putting AI-assisted engineering to work: internal coding and research models, running ahead of what’s publicly released, used by employees to speed up model development itself. For teams evaluating how fast capable coding agents might show up in commercial products, this is one of the few outside-audited numbers on how much lift labs are already getting internally, as opposed to marketing claims. Anthropic has since released Claude Fable 5.2 and Mythos 5.2, its newest public models.
How it compares
Outside commentators have compared the findings to AI 2027, an influential scenario written by former OpenAI researcher Daniel Kokotajlo’s AI Futures Project that forecast a key self-improvement breakthrough in early 2027. Marius Hobbhahn, CEO of Apollo Research, told The Verge the report suggests Anthropic’s pace is running faster than that scenario predicted. Separately, OpenAI said in September it is “making strong progress” toward an AI researcher that could contribute to deep learning and alignment work by March 2028.