Benchmark
GoBench
Elo is relative to this 9x9 Go KataGo ladder, with the random player fixed at zero; it is not a percentage or a human Go rank. Track 1 uses APIs without tools, Track 2 evaluates agent-built engines, and Track 3 evaluates tutorial-assisted API play. Context mode, reasoning effort, learning time, engine time limit, and the ladder snapshot affect comparisons. The cited LLM tables have 14-30 games per API configuration or 300 per engine, with 95% confidence intervals reported in the source. Current Track 2 uses a 60-second engine clock, unlike the older agent-in-the-loop paper results.
- Released
- 2026-09-15
- Evaluates
- reasoning
- Openness
- unknown
- Importer
- Model reports
GoBench links
GoBench results and reported scores
Reported benchmark scores · Model reports
Reported scores: 47
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | 2122.0 | 2026-10-19 | evidence |
| Claude Fable 5.1 | Anthropic | 2269.0 | 2026-10-19 | evidence |
| Claude Opus 5 | Anthropic | 2076.0 | 2026-10-19 | evidence |
| Claude Opus 5 | Anthropic | 2256.0 | 2026-10-19 | evidence |
| Claude Opus 5.5 | Anthropic | 2276.0 | 2026-10-19 | evidence |
| Claude Opus 5.5 | Anthropic | 3211.0 | 2026-10-19 | evidence |
| Claude Opus 5.5 | Anthropic | 3389.0 | 2026-10-19 | evidence |
| Claude Sonnet 5.5 | Anthropic | 2087.0 | 2026-10-19 | evidence |
| DeepSeek-V4-Flash-0731 | DeepSeek | 1189.0 | 2026-10-19 | evidence |
| DeepSeek-V4-Flash-0731 | DeepSeek | 1449.0 | 2026-10-19 | evidence |
| DeepSeek-V4.1-Flash | DeepSeek | 1381.0 | 2026-10-19 | evidence |
| DeepSeek-V4.1-Flash | DeepSeek | 1707.0 | 2026-10-19 | evidence |
| DeepSeek-V4.1-Flash-1017 | DeepSeek | 1634.0 | 2026-10-19 | evidence |
| Gemini 3.1 Pro | 1797.0 | 2026-10-19 | evidence | |
| Gemini 3.6 Flash | 1336.0 | 2026-10-19 | evidence | |
| Gemini 3.8 Flash | 1616.0 | 2026-10-19 | evidence | |
| gemini-4-argon | 2353.0 | 2026-10-19 | evidence | |
| GPT-5.6 Luna | OpenAI | 1091.0 | 2026-10-19 | evidence |
| GPT-5.6 Luna | OpenAI | 1058.0 | 2026-10-19 | evidence |
| GPT-5.6 Sol | OpenAI | 1846.0 | 2026-10-19 | evidence |
| GPT-5.6 Sol | OpenAI | 1929.0 | 2026-10-19 | evidence |
| GPT-5.6 Sol | OpenAI | 2609.0 | 2026-10-19 | evidence |
| GPT-5.6 Sol | OpenAI | 2472.0 | 2026-10-19 | evidence |
| GPT-5.6 Sol | OpenAI | 2388.0 | 2026-10-19 | evidence |
| GPT-5.6 Sol | OpenAI | 2492.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2227.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2301.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2568.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2011.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 3063.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 3230.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 3313.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 3264.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2341.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2550.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2386.0 | 2026-10-19 | evidence |
| GPT-6 Astra | OpenAI | 2375.0 | 2026-10-19 | evidence |
| GPT-6 Luna | OpenAI | 830.0 | 2026-10-19 | evidence |
| GPT-6 Luna | OpenAI | 1120.0 | 2026-10-19 | evidence |
| GPT-6 Sol | OpenAI | 1499.0 | 2026-10-19 | evidence |
| GPT-6 Sol | OpenAI | 2006.0 | 2026-10-19 | evidence |
| GPT-6.1 Sol | OpenAI | 2172.0 | 2026-10-19 | evidence |
| GPT-6.1 Sol | OpenAI | 2379.0 | 2026-10-19 | evidence |
| Grok 4.6 | xAI | 916.0 | 2026-10-19 | evidence |
| Grok 4.6 | xAI | 1045.0 | 2026-10-19 | evidence |
| Muse Spark 1.3 Contributor | Meta | 1449.0 | 2026-10-19 | evidence |
| Muse Spark 1.3 Contributor | Meta | 1415.0 | 2026-10-19 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite GoBench?
- GoBench Track 1-3 LLM comparison tables (2026-10-19) Roland Gao, 2026-10-19
- GoBench leaderboard and current track definitions Roland Gao