This scoreboard answers one question: how reliably does a model operate the Cast harness? It is not a general intelligence ranking. A model is certified once it clears 80% of the full behavior suite across exactly three fresh attempts per case, where every attempt on a case must agree (a case that only passes sometimes doesn't count). Core/Chain break the same score down by single-turn tool contracts vs. multi-turn stateful workflows (see Behavior Evals). Tokens are per-attempt averages; the time and turns columns are avg/median/p75/p95/p99 over every individual attempt. See Eval Methodology for the full methodology.

Per-signal breakdown