The question of whether AI is speeding up its own development has been argued mostly from outside the labs. METR, an independent AI safety research group, has now looked from inside one. After an 11-week investigation at Anthropic, it reports that the company's models are improving roughly twice as fast as a year ago, and that most of that acceleration comes from AI itself.
According to The Verge's account of the report, benchmark scores suggest Anthropic made about two years of historical progress over the past year. METR says it is confident this is not simply more training compute, which has driven most AI progress so far; as of October, the majority of the speed-up "is due to AI", and that share is growing. Staff using Anthropic's most capable internal model self-reported being about 12 times as productive as without AI help, roughly double what they reported a few months earlier using Mythos 5. Most of the gain still comes from AI assisting human researchers, though METR sees signs that fully autonomous contributions will soon matter more, and it found no clear bottlenecks to faster progress. The report also says Anthropic has finished training three unreleased models in the past six months, codenamed Model 2, Model 3 and Model 4.
The catch is what the public cannot see. Anthropic had the right to redact the final report and used it heavily. METR says it submitted a report of about 100 pages and that Anthropic removed a majority of the evidence METR found most informative, citing the risk of helping rivals; METR says most of what was cut was very high-level. As a result, METR writes, the public must rely on its judgement far more than would be ideal. Anthropic cofounder Sam McCandlish accepted that the redactions make the work harder to check.
The reaction has been sharp. Researchers quoted by The Verge, including Apollo Research's Marius Hobbhahn, say the pace looks faster than even the influential AI 2027 scenario, and critics point to the gap between Anthropic's calls for slower development and its own speed, as the company prepares to go public.
For users of Claude, the practical reading is that the models behind the tools in our reviews are likely to keep changing quickly, with stronger internal versions already in use. For everyone else, the open question is oversight. A self-reported productivity figure and a heavily redacted assessment are a start, not the kind of independent, checkable evidence that would settle how fast this is really going.