GameLogicBench: leaderboard
GameLogicBench (Che et al., 2026) asks coding agents to implement gameplay mechanics in Godot projects and judges each submission deterministically by checking game state at every simulation tick across hidden scenarios and seeds. Solve rate in percent over 72 tasks, one sample per model and scaffold (Claude Code, Codex CLI or OpenCode) at high effort with network egress sealed; higher is better.
Metric: Solve rate (%; all 72 tasks; 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 August 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (High) | 52.78 | #20 (Claude Opus 5) |
| 2 | Qwen 3.8 Max (High) | 44.44 | #55 (Qwen 3.8 Max) |
| 3 | GPT-5.6 Sol (High) | 41.67 | #29 (GPT-5.6 Sol) |
| 4 | DeepSeek V4 Pro (0813) (High) | 41.67 | #81 (DeepSeek V4 Pro (0813)) |
| 5 | GLM-5.2 (High) | 37.5 | #137 (GLM-5.2) |
| 6 | Kimi K3 (High) | 36.11 | #53 (Kimi K3) |
| 7 | DeepSeek V4 Flash (0731) (High) | 31.94 | #136 (DeepSeek V4 Flash (0731)) |
| 8 | GPT-5.6 Terra (High) | 29.17 | #60 (GPT-5.6 Terra) |
| 9 | Claude Sonnet 5 (High) | 25 | #108 (Claude Sonnet 5) |
| 10 | DeepSeek V4 Pro (High) | 25 | #181 (DeepSeek V4 Pro) |
| 11 | MiniMax-M3 (High) | 19.44 | #265 (MiniMax-M3) |
| 12 | DeepSeek V4 Flash (High) | 16.67 | #254 (DeepSeek V4 Flash) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gamelogicbench · How It Works · Data refreshed daily, snapshot 2026-11-12.