LMGame-Bench Sokoban: leaderboard
GamingAgent pushes boxes onto goals until deadlock; native leaderboard boxes-on-target score. Primary GamingAgent harnessed cohort, distinct from model-only tab; full published score precision and source run annotations retained.
Metric: Score. Source: huggingface.co. Saturation forecast: Around December 2026. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (High) | 11 | #98 (GPT-5) |
| 2 | O3 (2025-04-16) | 8 | #121 |
| 3 | Grok 4 | 6.33 | #181 |
| 4 | Grok 3 Mini Beta | 5.67 | #296 |
| 5 | O4 Mini (2025-04-16) | 5.33 | #179 |
| 6 | DeepSeek R1 0528 | 4.67 | #222 |
| 7 | Gemini 2.5 Pro (Preview 05-06) | 4.33 | #147 |
| 8 | GPT-OSS-120B | 4 | #341 |
| 9 | GPT-OSS-20B | 4 | #518 |
| 10 | Claude Opus 4 (20250514) | 4 | #154 |
| 11 | Claude Sonnet 4 (20250514) | 3 | #213 |
| 12 | Claude 3.7 Sonnet (20250219) | 2.33 | #201 |
| 13 | O1 (2024-12-17) | 2.33 | #149 |
| 14 | Qwen 3 235B A22B FP8 | 2.33 | #301 |
| 15 | Gemini 2.5 Flash (Preview 04-17) | 1.67 | #237 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=lmgame-bench-sokoban · How It Works · Data refreshed daily, snapshot 2026-10-20.