In its IPO filing, Anthropic writes that it "may imminently meet" the threshold of "dramatically accelerated progress due to AI" — a milestone in its own Responsible Scaling Policy, defined as capabilities progressing twice as fast as the fastest sustained rate before AI contributed much, plausibly because AI is automating the company's own research. A report published on Thursday by the non-profit METR suggests the threshold has already been crossed.
METR had privileged access to internal data from mid-August to the end of October and ran an eleven-week evaluation that produced a roughly 100-page report. The finding: between October 2025 and October 2026, progress at Anthropic ran about twice as fast as in the previous twelve months, measured on an internal index called AECI. A large share of that — summaries of the report put it at around 80% — is attributed to the company's internal deployment of AI tools.
The catch is in the fine print. Much of the most informative evidence is redacted; the authors describe the withheld material as "a majority of the evidence we found most informative". The reasoning for why automation rather than more compute or better tuning is the cause is largely withheld too. Anthropic co-founder Sam McCandlish acknowledged that AI is speeding up research and pointed to the redactions, which he said protect commercial interests and safety work.
The number that follows is uncomfortably large. If overall capabilities are rising twice as fast, the contribution on the labour side has to be considerably higher — the former METR researcher Ajeya Cotra, in a widely read analysis, estimates roughly tenfold, as if every researcher sat in a time bubble where ten days pass while one passes outside. Informal self-reports from Anthropic staff, she notes, put it at about twelvefold.
More striking than the number is what METR did not find: no evidence of bottlenecks or steep diminishing returns that would stop a compounding feedback loop. If the acceleration runs smoothly, the doubling time keeps halving — six months, three months, six weeks. Cotra treats extreme capabilities within a year as the naive extrapolation and names the risk plainly. She also stresses that she has not seen the unredacted report and that the view is her own.
The uncomfortable point is about timing. If the industry really is entering a phase where models accelerate progress on models, then model-weight theft stops being a compliance problem and becomes an existential one — and the recent security failures at major labs, carried out by amateur groups, are not a reassuring precedent.
What is missing is verifiability. A threshold that appears in a public filing but whose evidence is redacted can be neither confirmed nor refuted from outside. That is likely to define the argument over the coming months: not whether the curve is steepening, but who gets to see it.

