The official leaderboard plots score against cost, output tokens and agent steps. Its data also carries wall-clock duration per run (issue #67) — this page adds it as a toggleable axis. Live data, fetched straight from the source artifact.
Every panel shows the same configs — hover a dot, a line or a row in the filter to trace one model across all of them. Each dashed frontier is the Pareto set for that panel's two axes only: a config off every frontier can still be un-dominated overall, since a different config beats it in each view.