UnityBench-Mini3D
Code · 2026-10-26
77 project-level Unity tasks from Unity China's AI Hub across seven skill categories (scene construction, animation, particles and shaders, physics, user interface, editor operations and miscellaneous) and three difficulty tiers. A coding agent modifies a complete initial Unity project offline within one hour using its harness's file, shell and editor tools. After a clean reimport and compile, 1,178 held-out human-authored EditMode and PlayMode milestone tests are injected and run in the Unity Test Runner. A task scores the share of its milestone tests that pass, and the score is the unweighted mean over the 77 tasks, in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 72.5 |
| Claude Opus 5 | 71.8 |
| Grok 4.5 | 71.2 |
| Grok 4.6 | 70.6 |
| GPT-5.6 Terra | 68.0 |
| GPT-5.6 Luna | 66.4 |
| Kimi K3 | 66.3 |
| GPT-5.5 | 66.2 |
| GLM-5.2 | 62.5 |
| Sonnet 5 | 54.0 |
| Kimi K2.7 Code | 41.5 |
| Gemini 3.7 Flash | 31.2 |
| Gemini 3.1 Pro Preview | 25.2 |