Learn2Play Bench — Mean Episode Score
Games · 2026-10-06
Learn2Play Bench's Mean metric: the normalized terminal score averaged over every evaluated episode (5 per normal game, 10 per challenge game) and three trials, then across the 20 games with challenge games weighted 3. Early, pre-learning episodes count in full, so it measures consistency throughout evaluation rather than peak play. A diagnostic beside the Max primary.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 62.3 |
| Claude Opus 5.5 | 61.0 |
| GPT-6 Astra | 56.1 |
| GPT-5.6 Sol | 54.8 |
| Gemini 3.5 Flash | 52.2 |
| Kimi K3 | 50.2 |
| Gemini 3.7 Flash | 39.9 |
| Sonnet 5 | 35.6 |
| GPT-5 Mini | 22.3 |