gg-bench: leaderboard
Zero-sum grid-based game competition benchmark from UC Berkeley. Evaluates domain-general capabilities and strategic reasoning without relying on human-curated QA tasks.
Metric: Mean win rate vs RL opponents (%). Source: github.com. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 36.28 | #183 |
| 2 | DeepSeek R1 | 32.5 | #251 |
| 3 | O3 Mini | 31.08 | #272 |
| 4 | Claude 3.7 Sonnet | 9.53 | #249 |
| 5 | GPT-4o | 8.94 | #343 |
| 6 | GPT-4o Mini | 7.64 | #599 |
| 7 | Llama 3.3 70B Instruct | 7.42 | #534 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=gg-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.