Scored per capability and per domain, from recognised public benchmarks (Humanity's Last Exam, GPQA Diamond, FrontierMath, Terminal-Bench, OSWorld 2.0, SimpleQA Verified…) and from real governed agent runs. Every number traces to the benchmark, who measured it and when. There is no single winner: pick per intent.
Dot colour = where the number comes from: cited benchmarks · real governed runs. Only measured numbers are shown here; a dash means no published result yet. Cell opacity tracks confidence.
Method. Per benchmark, a model's result is expressed as a share of the best published result, since benchmarks differ in difficulty. Shares are averaged per capability, weighted by benchmark status (frontier benchmarks count fully, near-saturated ones half, saturated ones are not scored) and by who measured the result (independent runs count more than leaderboard entries, which count more than vendor claims). Two scores within 5 points are within measurement noise.
Benchmark data: Epoch AI, ‘Capabilities & benchmarking’, published online at epoch.ai, licensed CC BY 4.0. Results Epoch collected from other sources are credited to those sources.