A common scale, with uncertainty

ECI fits a capability scale across overlapping benchmark results. It is useful for following broad progress, while individual tests reveal what a model can actually do. ECI points are not percentages. Epoch’s difficulty index, EDI, describes benchmarks and stays separate from model leaderboards.

Beyond a single index

The collection includes FrontierMath and its distinct releases, open mathematical problems, MirrorCode, Earthborne Rangers, Furniture Assembly, puzzle evaluations and externally collected benchmark results. The table sources identify the original publisher as well as Epoch’s data snapshot.