AI benchmarks

A benchmark is a fixed set of tasks with a scoring rule, so it shows how models compare on that kind of work and nothing more. The table below lists the benchmarks we track, what each one measures, its metric and how many model results we have recorded. Each benchmark page explains the method and its caveats, and lists every reported result with the setting it was measured under (reasoning effort, tools, agent harness or leaderboard view), who reported it and the source. Scores move with the setting, so we compare them only within the same setting, and results a provider publishes about its own models are labeled self-reported.

The last column shows which of our rankings use a benchmark as a criterion. How we rank models explains how scores, prices and capabilities are weighted. To read results with their settings and sources, and to judge whether a gap of a few points matters, see how to compare AI models fairly. How to choose an AI model shows where benchmarks fit next to price, context window, licence and latency, and reasoning models explained explains the reasoning effort setting that many results are measured under.