A nonprofit reviewer says Anthropic's AI models now improve about twice as fast, and AI doing its own research is likely the main reason. Here is what that could mean.
Anthropic's AI models are getting better about twice as fast as before October 2025, and most of the speedup appears to come from Anthropic using its own AI to help do the research. That is the finding of METR, a nonprofit that evaluates AI risks, in an assessment published on November 5.
METR spent about eleven weeks, from August 12 to October 29, inside Anthropic's internal research data. It interviewed around 15 staff and reviewed benchmark results, headcount, training compute, surveys and lines of code. Its main measure is an internal score that combines results from many tests. That score rose from about 15 points a year to 29, a 1.9x speedup. The likely range is 1.5x to 2.8x.
The score tracks test results, not research output directly. Staff polls put self-reported productivity gains from Anthropic's most capable internal model at about 12x, but METR called that only very loose evidence. One researcher put AI's share of the work at 15 to 20%, according to The Decoder. METR also said AI cannot yet make highly significant, fully autonomous contributions to the most important parts of building models.
Anthropic's own safety policy has a threshold for doubling the rate of progress. Its public August report said the speedup was under 2x. METR calculated that a hidden figure in the private version implies 1.97. It called it plausible that the line was crossed by July 15, though METR and Anthropic staff lean toward it not having been crossed. On November 2, Anthropic's amended S-1 filing said it may imminently meet its definition of dramatically accelerated progress attributable to AI.
Anthropic had the right to redact information. METR said most of the evidence it found most useful was hidden, so the public must rely on METR's judgment more than is ideal. Anthropic also changed the project midway to drop a check on how it splits computing power between capabilities and safety research. METR warned that progress could get even faster over the coming year.
If AI keeps speeding up its own development, the tools people use at work and at home may change faster than they can plan for.
A neutral reviewer says Anthropic's AI is improving about twice as fast, largely thanks to AI, so expect the tools you use to change quickly.
World models, AI that learns physics from video, are being built into home robots and driverless cars, and your own videos may help train them.
Read more: AI that learns how the physical world works is coming to robots and carsA new Alibaba-led test of real coding requests found the best AI agent fully finished just 26.4 percent of tasks, so human review still matters.
Read more: AI coding helpers fully finish only about 1 in 4 real tasksA University of Chicago law professor asked Claude to fix page numbers. It deleted his syllabus and a year of notes instead. Dropbox saved him, not the AI.
Read more: An AI assistant deleted a law professor's year of lecture notes
