Kagi LLM Chess Puzzles: leaderboard
Kagi LLM chess puzzles: each model is given 1,000 random Lichess puzzles as FEN board states and asked for the best move. Score is the percentage of puzzles solved (0–100, higher is better); illegal-move counts and an adjusted Elo are also published upstream. Results were last updated in April 2025 and cover mostly 2024-era models.
Metric: Puzzles Solved (%). Source: github.com. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.5 (Preview) | 84.8 | #260 |
| 2 | GPT-4.1 | 57.9 | #258 |
| 3 | GPT-4o | 53.6 | #357 |
| 4 | GPT-4 Turbo (Preview) | 22.9 | #416 |
| 5 | GPT-4 | 19.5 | #459 |
| 6 | Claude 3 Opus (20240229) | 7.2 | #428 |
| 7 | Claude 3 Haiku (20240307) | 3.8 | #780 |
| 8 | Claude 3 Sonnet (20240229) | 2.3 | #694 |
| 9 | GPT-3.5 Turbo | 2.3 | #858 |
| 10 | Claude Instant 1.2 | 1 | #1003 |
| 11 | Mixtral 8x7B Instruct | 0.9 | #1085 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kagi-llm-chess-puzzles · How It Works · Data refreshed daily, snapshot 2026-11-05.