GameLogicBench - Repo: leaderboard
The Repo tier of GameLogicBench (Che et al., 2026): 23 tasks that implement a mechanic in place inside a real open-source Godot project whose other systems must keep their invariants. Solve rate in percent from tick-level state assertions, one sample per model and scaffold at high effort with network egress sealed; higher is better.
Metric: Solve rate (%; the 23 Repo tasks (a mechanic inside a real open-source project); one sample per model-scaffold configuration with effort=high, a 3,600 s limit and network egress sealed; a task counts as solved only if the submitted Godot 4.4 project passes every tick-level state assertion in every judge-selected scenario and seed). Source: arxiv.org. Saturation forecast: Around December 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (High) | 43.48 | #20 (Claude Opus 5) |
| 2 | GPT-5.6 Sol (High) | 34.78 | #29 (GPT-5.6 Sol) |
| 3 | DeepSeek V4 Pro (0813) (High) | 34.78 | #81 (DeepSeek V4 Pro (0813)) |
| 4 | GLM-5.2 (High) | 30.43 | #137 (GLM-5.2) |
| 5 | DeepSeek V4 Flash (0731) (High) | 26.09 | #136 (DeepSeek V4 Flash (0731)) |
| 6 | GPT-5.6 Terra (High) | 21.74 | #60 (GPT-5.6 Terra) |
| 7 | Qwen 3.8 Max (High) | 21.74 | #55 (Qwen 3.8 Max) |
| 8 | Kimi K3 (High) | 13.04 | #53 (Kimi K3) |
| 9 | DeepSeek V4 Pro (High) | 13.04 | #181 (DeepSeek V4 Pro) |
| 10 | DeepSeek V4 Flash (High) | 13.04 | #254 (DeepSeek V4 Flash) |
| 11 | MiniMax-M3 (High) | 13.04 | #265 (MiniMax-M3) |
| 12 | Claude Sonnet 5 (High) | 8.7 | #108 (Claude Sonnet 5) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gamelogicbench-repo · How It Works · Data refreshed daily, snapshot 2026-11-12.