CADWorld: leaderboard
CADWorld (Dong et al., 2026) is a computer-use benchmark of 200 long-horizon tasks in the FreeCAD mechanical-design application, spanning sketching, part modeling, assembly, CAM, FEM, measurement, mesh processing and technical drawing. Agents work from screenshots with mouse and keyboard actions for up to 100 steps, and a task counts only when every executable check on the saved CAD project passes; success rate in percent, higher is better.
Metric: Task success rate (%; 200 long-horizon FreeCAD tasks in 11 workflow categories; the agent acts on 1920x1080 screenshots through GUI actions for at most 100 steps; success only when every task-specific executable check on the saved .FCStd artifact and auxiliary outputs passes). Source: arxiv.org. Saturation forecast: Around October 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 17.5 | #121 (GPT-5.4) |
| 2 | Claude Opus 4.8 (Medium) | 16 | #56 (Claude Opus 4.8) |
| 3 | Kimi K2.6 (Non-reasoning) | 7.5 | #173 (Kimi K2.6) |
| 4 | Qwen 3.6 35B A3B (Non-reasoning) | 0 | #461 (Qwen 3.6 35B A3B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cadworld · How It Works · Data refreshed daily, snapshot 2026-11-12.