Compare what each model's ECI rests on

A model's ECI is, to within a few tenths of a point, a weighted average of the ECIs its individual benchmark scores imply (dots). Weights depend on how steep each benchmark's curve is where the model sits and how widely its scores scatter around the curve; flat, saturated or widely scattered benchmarks count less. Both menus list frontier models and near-frontier models released since GPT-5 (within 8 ECI points of the frontier at release; the number in parentheses is how far behind).

Compare with

Hover over or tap a dot for the score behind it; click a benchmark name for Epoch's page on it.

† Epoch rescales this benchmark's scores before fitting: random guessing counts as 0% and, for some, a maximum below 100% counts as 100%. Scores here are the rescaled ones; hover a dot for the raw score. Data: Epoch AI Benchmarking Hub, ECI as published (data as of Nov 17, 2026).