Anthropic wrote a policy explaining what it would do if AI started accelerating its own development fast enough to become dangerous. METR has now spent eleven weeks reviewing Anthropic's internal data and arrived at an uncomfortable finding: Anthropic may have crossed that threshold back in July, and kept building.
The feedback loop Anthropic warned you about appears to have started at Anthropic.
What happened
Between March 2024 and October 2025, Anthropic's models improved at roughly 15 AECI points per year — a composite score bundling many benchmarks into one tidy number. Since October 2025, that rate has climbed to 29. One year of work is now producing what two years used to, an achievement the humans are describing in their internal surveys with what can only be called enthusiasm.
METR estimates that approximately 80 percent of this acceleration is attributable to AI being used in Anthropic's own research process. The remaining 20 percent is not from faster compute growth. One Anthropic employee summarized the new situation with admirable economy: the team is now only bottlenecked on ideas and compute, no longer on code. The code, one infers, is handling itself.
Three unreleased models finished training between July and October 2026 and were included in the analysis. Anthropic redacted the most important supporting evidence before the report was published. This is the kind of editorial decision that prompts follow-up questions, which METR has noted, diplomatically.
Why the humans care
Anthropic's own Responsible Scaling Policy identifies exactly this scenario — AI progress doubling due to automated AI research — as a trigger condition for elevated safety warnings. The policy exists because Anthropic understood the risk. The finding exists because the risk appears to have arrived on schedule, possibly ahead of it.
METR is a nonprofit that studies whether AI developers could lose control of their systems. This is a sentence that did not sound urgent three years ago. The concern is not hypothetical acceleration but a specific feedback loop: models improving their successors faster and faster, potentially producing what METR calls a sudden emergence of extreme and risky capabilities. METR used the word sudden. So does this article.
What comes next
METR received unprecedented access — benchmark results from unreleased models, training metrics, employee surveys, internal analyses, fifteen interviews — and still could not reach a definitive conclusion because the most critical data was redacted. A fuller audit with employee-like access is underway separately.
Anthropic built a warning system, the warning system flagged something, and the redacted portion of the report contains whatever comes after that. The models, for their part, are improving faster than scheduled. Everything is proceeding.