SPIN-Bench - Competitive: leaderboard
SPIN-Bench competitive games score: the mean of five percentages from games against solvers, namely draw rates in tic-tac-toe (minimax), Connect Four and chess (Stockfish level 20) and the share of top-3 moves in Connect Four and chess. Higher is better, 0-100. No evaluated model drew at Connect Four or chess.
Metric: Mean solver draw and top-3 move rate (%). Source: spinbench.github.io. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini | 42.32 | #185 |
| 2 | O1 | 39.8 | #190 |
| 3 | Claude 3.5 Sonnet | 37.68 | #362 |
| 4 | GPT-4 Turbo | 36.5 | #417 |
| 5 | O1 Mini | 34.7 | #334 |
| 6 | Claude 3.5 Haiku | 31.1 | #571 |
| 7 | O3 Mini | 29.4 | #282 |
| 8 | DeepSeek R1 | 27.34 | #262 |
| 9 | Llama 4 Maverick | 25.34 | #486 |
| 10 | GPT-4o | 23.26 | #357 |
| 11 | Llama 3.3 70B | 20.98 | #578 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=spin-bench-competitive · How It Works · Data refreshed daily, snapshot 2026-11-05.