Benchmarks
331 sources · 1,258 models × 2,980 benchmarks · 155,643 scoresWorkflowArena (Navers Lab, Einsia.AI and Tsinghua University) tests workflow orchestration: whether an agent can decide if and how to split long-horizon work across worker agents it launches itself. Its 145 tasks span 8 domains (software engineering, enterprise operations, science, formal mathematics in Lean, finance, law, video and document content, and 3D production) and 17 task forms built from real repositories, filings, agreements and media. Each task gives a workspace and an integrated deliverable whose parts depend on one another, assessed by a hidden verifier. The agent has its harness's tools plus a Workflow tool that runs up to 16 concurrent workers (fresh instances of the same model, harness and effort), native sub-agents are disabled so every delegation is recorded, and the main agent and its workers share a four-hour wall-clock budget. A task is resolved only when its deliverable meets every acceptance criterion of its domain; Overall Resolved, the headline, is the equal-weight mean of the eight domain resolved rates. Scores are for a model in a harness (Claude Code or Codex), not for the model alone.
No signal or difficulty estimate. Too few models have been measured on this benchmark to estimate its difficulty.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | max · Claude Code | 37.0 | |
| 2 | Claude Sonnet 5.5 | max · Claude Code | 33.2 | |
| 3 | GPT-6 Astra | x-high · Codex | 30.8 | |
| 4 | GPT-6.1 Sol | x-high · Codex | 29.5 | |
| 5 | Kimi K3.1 Pro | max · Claude Code | 16.5 | |
| 6 | GPT-6 Luna | max · Codex | 13.3 | |
| 7 | GLM-5.3 Flash | max · Claude Code | 4.4 |