Learn2Play Bench - Max: leaderboard
LLM agents play 20 unfamiliar games with hidden rules (breeding butterflies toward a target, running an inn with mysterious constraints, synthesizing compounds) over repeated episodes and must improve from experience alone, without weight updates. Max is the average peak score across three trials, normalized to each game's reference ceiling with challenge games weighted 3:1; every row is a backbone in the same OpenCode harness. A peak can come from any scored episode, so Max alone does not show how much a model learned.
Metric: Normalized Max (% of each game's reference ceiling, challenge games weighted 3). Source: liushiliushi.github.io. Saturation forecast: Around February 2028. 11 models tracked. Updated 19 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 80.1 | #9 | |
| 2 | GPT-6 Astra | Medium | 76.2 | #10 |
| 3 | Claude Opus 5 | 74.5 | #26 | |
| 4 | Qwen 3.8 Max | 72.5 | #66 | |
| 5 | Gemini 3.5 Flash | 72 | #108 | |
| 6 | Kimi K3 | 68.5 | #61 | |
| 7 | Gemini 3.7 Flash | 58.3 | #42 | |
| 8 | GPT-5.6 Sol | 57 | #36 | |
| 9 | Claude Sonnet 5 | 55.9 | #118 | |
| 10 | DeepSeek V4 Pro | 54.2 | #198 | |
| 11 | GPT-5 Mini | 34 | #365 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=learn2play-bench-max · How It Works · Data refreshed daily, snapshot 2026-11-20.