LisanBench — leaderboard
LisanBench evaluates LLMs on open-ended word-chain generation, stressing vocabulary, planning, recall, persistence, and exact edit-distance rule following.
Metric: Mean Path Length / Current Maximum. Source: lisanbench.com. Status: saturated. 150 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.7 (xHigh) | 1 |
| 2 | GPT-5.5 (Medium) | 0.68 |
| 3 | Claude Opus 4.8 (High) | 0.66 |
| 4 | Gemini 3.1 Pro (Preview) (High) | 0.45 |
| 5 | Claude Sonnet 5 (High) | 0.4 |
| 6 | GPT-5.4 (Medium) | 0.37 |
| 7 | GPT-5.6 Sol (Medium) | 0.33 |
| 8 | Gemini 3 Pro (Preview) (High) | 0.32 |
| 9 | Gemini 3.5 Flash (High) | 0.29 |
| 10 | GPT-5.2 (Medium) | 0.25 |
| 11 | Gemini 3.1 Pro (Preview) (Low) | 0.24 |
| 12 | O3 (Medium) | 0.23 |
| 13 | GPT-5.6 Terra (Medium) | 0.23 |
| 14 | DeepSeek V4 Pro (High) | 0.23 |
| 15 | DeepSeek V3.2 (Thinking) | 0.16 |
Interactive version: theaggregate.ai/benchmark?slug=lisanbench · How the rankings work · Data refreshed daily, snapshot 2026-07-22.