METR finds AI R&D capabilities improving faster than benchmarks capture
The non-obvious implication is that headline capability metrics based on time horizons may systematically understate what frontier models can do on real research work, which affects any forecast built on those numbers. Teams building internal evaluation suites or making procurement decisions can use METR's new framing to stress-test how they compare models. Readers tracking US lab progress against Chinese rivals should note this is a US measurement effort whose results will feed capability comparisons across both ecosystems.