SnorkelSpatial: leaderboard
Snorkel AI's archived SnorkelSpatial benchmark: 330 text-only grid problems in which a board and particles are moved and rotated through 10 to 200 actions before questions about final positions and orientations. Accuracy@1 against programmatic ground truth, with up to 10 attempts for valid JSON; %, higher is better. Frozen roster.
Metric: Accuracy@1 (%, 330 spatial reasoning questions). Source: snorkel.ai. Saturation forecast: Estimated already saturated. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 99 | #83 |
| 2 | Grok 4 Fast (Reasoning) | 84.85 | #266 (Grok 4 Fast) |
| 3 | O3 | 76.67 | #132 |
| 4 | GPT-5 | 73.94 | #105 |
| 5 | GPT-OSS-120B | 52.73 | #346 |
| 6 | GPT-5 Mini | 45.45 | #193 |
| 7 | Claude Opus 4.1 | 45.15 | #146 |
| 8 | Magistral Medium 1.2 | 44.24 | #444 |
| 9 | Claude Opus 4 | 40.3 | #170 |
| 10 | O3 Mini | 37.88 | #282 |
| 11 | Claude Sonnet 4 | 33.33 | #207 |
| 12 | GPT-5 Nano | 26.67 | #439 |
| 13 | Claude 3.7 Sonnet | 21.52 | #259 |
| 14 | Gemini 2.5 Flash | 18.79 | #255 |
| 15 | Llama 4 Scout | 15.45 | #648 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=snorkelspatial · How It Works · Data refreshed daily, snapshot 2026-11-05.