LLM Chess (Saplin): leaderboard
LLMs play Black through an AutoGen conversational chess harness against Random and Komodo Dragon opponents, using actions to inspect boards/legal moves and execute UCI moves. Native ratings solve the mixed-opponent Elo likelihood with Dragon anchors of 125×(level+1), calibrated Random opponents and source 35-point White-advantage adjustment. Signed negative ratings are legitimate and unbounded. Retain native three-decimal Elo, opponent counts and confidence margins. 55 source rows with mathematically undefined/unsupported anchored estimates stay blank rather than zero. Evaluated model suffixes, effort and quantization labels are preserved; no broad checkpoint remapping.
Metric: Native anchored-opponent Elo (signed points). Source: maxim-saplin.github.io. Saturation forecast: Around November 2027. 178 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (High) | 1575.25 | #3 (Claude Opus 5.5) |
| 2 | GPT-5.6 Sol (xHigh) | 1549.71 | #19 (GPT-5.6 Sol) |
| 3 | GPT-5.6 Sol (High) | 1548.47 | #19 (GPT-5.6 Sol) |
| 4 | Gemini 3.8 Flash (High) | 1546.08 | #21 (Gemini 3.8 Flash) |
| 5 | GPT-5.5 (Medium) | 1532.16 | #31 (GPT-5.5) |
| 6 | GPT-5.5 (High) | 1528.44 | #31 (GPT-5.5) |
| 7 | Gemini 3.8 Flash (Medium) | 1526.51 | #21 (Gemini 3.8 Flash) |
| 8 | Gemini 3.1 Pro (Preview) (High) | 1511.39 | #58 (Gemini 3.1 Pro (Preview)) |
| 9 | GPT-5.4 (High) | 1322.3 | #79 (GPT-5.4) |
| 10 | GPT-5.6 Terra (xHigh) | 1293.01 | #41 (GPT-5.6 Terra) |
| 11 | GPT-5.4 (Medium) | 1252.93 | #79 (GPT-5.4) |
| 12 | GPT-5.6 Sol (Medium) | 1239.97 | #19 (GPT-5.6 Sol) |
| 13 | GPT-5.6 Terra (High) | 1210.74 | #41 (GPT-5.6 Terra) |
| 14 | Claude Sonnet 5.5 (High) | 1194.57 | #11 (Claude Sonnet 5.5) |
| 15 | GPT-5.6 Luna (xHigh) | 1185.3 | #68 (GPT-5.6 Luna) |
No result here: #6 GPT-6 Astra, #7 GPT-6.1 Sol, #8 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=llm-chess-saplin · How It Works · Data refreshed daily, snapshot 2026-10-20.