Atlas

Benchmarks

← All benchmarks

UnityBench-Mini3D (Medium)

Code · 2026-10-26

The 29 Medium-tier UnityBench-Mini3D tasks. A coding agent modifies a complete initial Unity project offline within one hour, and held-out human-authored EditMode and PlayMode milestone tests run in the Unity Test Runner on the final project. The score is the mean over these tasks of the share of each task's milestone tests that pass, in percent.

Top models (higher is better)

ModelScore
Claude Opus 574.3
Grok 4.570.9
Kimi K368.6
GPT-5.6 Sol67.0
Grok 4.666.1
GPT-5.6 Luna66.0
GPT-5.6 Terra65.4
GLM-5.263.4
GPT-5.563.0
Sonnet 549.8
Kimi K2.7 Code41.8
Gemini 3.1 Pro Preview26.9
Gemini 3.7 Flash23.9
Loading Atlas data…