Epoch AI reports AI capabilities progress has sped up
A two-segment regression across 149 models found capability gains almost doubled in pace after April 2024, a break Epoch linked to the rise of reasoning models.
- Benchmarks & progress
- Minor
Epoch AI published an analysis finding that the pace of frontier AI capability progress had roughly doubled since April 2024. Using its Epoch Capabilities Index, fitted to 149 frontier models released between December 2021 and December 2025, the best-performing models gained about 8.3 index points a year before April 2024 and about 15.5 points a year after — a roughly 1.85-times acceleration, with a 90% confidence interval spanning 1.3 to 3.1 times. A two-segment regression fit the data better than a single trend line by Epoch’s own statistical test, and the researchers said a resampling check gave 90% confidence in the location of the breakpoint.
Epoch said the finding was corroborated by METR’s separate time-horizon benchmark, which showed a comparable acceleration of around 40% beginning in October 2024. It attributed the shift to the emergence of reasoning models and a growing emphasis on reinforcement learning among leading labs over the same period, rather than to a single model release. The report adds a second, independently derived data series to a debate — already carried by METR’s own doubling-time findings — over whether frontier progress accelerated, plateaued or simply continued on trend through 2025.