Kaggle OracleBench: leaderboard
Community benchmark on Kaggle by utilisateurrichard: eight tasks checked against optimal solvers: Othello, Connect 4 and Kuhn poker against optimal opponents, Mastermind, Nim with a confidence bet, misere Nim, grid navigation among distracting prompts and next-value sequence prediction. Each task scores 0-1 and an invalid move ends a game; mean over the eight as a percentage, higher is better.
Metric: Mean Game Score (%). Source: www.kaggle.com. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 59.38 | #90 |
| 2 | Gemini 3.1 Pro (Preview) | 56.11 | #64 |
| 3 | GLM-5 | 51.25 | #149 |
| 4 | Gemini 2.5 Pro | 50.86 | #159 |
| 5 | Claude Opus 4.6 | 49.37 | #67 |
| 6 | Gemini 2.5 Flash | 36.38 | #255 |
| 7 | GPT-OSS-120B | 33.93 | #346 |
| 8 | Claude Sonnet 4.6 | 30.14 | #94 |
| 9 | Gemini 3.1 Flash Lite (Preview) | 25.67 | #201 |
| 10 | Claude Haiku 4.5 | 22.42 | #287 |
| 11 | Qwen 3 235B A22B 2507 Instruct | 18.37 | #315 |
| 12 | Gemma 4 26B A4B | 13.25 | #360 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kaggle-oraclebench · How It Works · Data refreshed daily, snapshot 2026-11-05.