SPIN-Bench: leaderboard
SPIN-Bench (COLM 2025) average score: the mean of PDDL plan accuracy, draw rates against solvers in tic-tac-toe, Connect Four and chess, top-3 move rates in Connect Four and chess, and 2- to 5-player Hanabi scores converted to percent of the 25-point maximum. Higher is better, 0-100.
Metric: Average score (%). Source: spinbench.github.io. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 49.8 | #190 |
| 2 | O4 Mini | 45.7 | #185 |
| 3 | DeepSeek R1 | 36.6 | #262 |
| 4 | Claude 3.5 Sonnet | 34.3 | #362 |
| 5 | O3 Mini | 33.1 | #282 |
| 6 | O1 Mini | 33 | #334 |
| 7 | GPT-4 Turbo | 27.5 | #417 |
| 8 | Llama 4 Maverick | 20.9 | #486 |
| 9 | GPT-4o | 20.8 | #357 |
| 10 | Claude 3.5 Haiku | 20.8 | #571 |
| 11 | Llama 3.3 70B | 13.1 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-05.