AI may be starting to speed up the work that produces better AI. That possibility puts the research behind the models in the story, not just the products we see.
METR, an independent AI evaluator, used private data supplied by Anthropic to estimate that its models were improving on clearly defined tasks about twice as fast as in the earlier period. The estimate has substantial uncertainty, and it describes Anthropic’s internal models, not the whole industry.
If AI tools help researchers make progress faster, they could also make it easier to build the next generation of AI. That puts the pace of research itself on the safety agenda.
METR tracked performance on clearly defined tasks over time, a bit like timing runners on the same course. That shows how quickly performance improves, but it does not prove that models are independently building their successors.
In a separate review, METR called the evidence a possible early warning that AI-assisted research may be accelerating. It still judged direct catastrophic risk from the models it assessed to be low. The signal is worth monitoring, not a prediction of imminent disaster.
For AI labs and safety evaluators, the question is becoming not only what a model can do, but whether it helps people improve the next one. The open question is whether this pattern continues, and whether independent measurements can confirm it beyond Anthropic’s internal data.
