ComposioBENCH
Try Bench

Models

Every model on the bench, ranked by its best run on the same 30 stateful tool-use tasks. Success moves with the harness, so each model page shows every harness it has run through.

Every model on the bench
  1. 01GPT-6 AstraClineCodexHermes AgentCommand CodeClaude CodeOpenCodePi Agent
    79%
    Best time / task125s
  2. 02DeepSeek V4 ProClaude CodeCodexOpenCodeHermes AgentPi AgentDeepSeekCommand CodeCline
    60%
    Best time / task327s

Ranked by best success, then harnesses run · best-run figures may come from different harnesses · Composio Benchmark