GameLogicBench - Atom: leaderboard
The Atom tier of GameLogicBench (Che et al., 2026): 21 tasks that each isolate one gameplay mechanic in a minimal Godot game built for the benchmark, judged by tick-level state assertions across hidden scenarios. Solve rate in percent, one sample per model and scaffold at high effort; higher is better.
Metric: Solve rate (%; the 21 Atom tasks (one mechanic in a minimal game); 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 February 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (High) | 61.9 | #20 (Claude Opus 5) |
| 2 | Qwen 3.8 Max (High) | 61.9 | #55 (Qwen 3.8 Max) |
| 3 | GPT-5.6 Sol (High) | 57.14 | #29 (GPT-5.6 Sol) |
| 4 | Kimi K3 (High) | 52.38 | #53 (Kimi K3) |
| 5 | GLM-5.2 (High) | 52.38 | #137 (GLM-5.2) |
| 6 | DeepSeek V4 Pro (0813) (High) | 52.38 | #81 (DeepSeek V4 Pro (0813)) |
| 7 | Claude Sonnet 5 (High) | 47.62 | #108 (Claude Sonnet 5) |
| 8 | GPT-5.6 Terra (High) | 42.86 | #60 (GPT-5.6 Terra) |
| 9 | DeepSeek V4 Flash (0731) (High) | 38.1 | #136 (DeepSeek V4 Flash (0731)) |
| 10 | DeepSeek V4 Pro (High) | 33.33 | #181 (DeepSeek V4 Pro) |
| 11 | MiniMax-M3 (High) | 28.57 | #265 (MiniMax-M3) |
| 12 | DeepSeek V4 Flash (High) | 23.81 | #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-atom · How It Works · Data refreshed daily, snapshot 2026-11-12.