BoardBench

Go

19×19 Go, Chinese rules, komi 7.5, given as SGF. KataGo scores every legal move and pass by win rate; the answer is a GTP coordinate. See Methodology.

  1. Claude Opus 5.5High67.8$0.03
  2. Claude Haiku 5.5High66.7<$0.01
  3. GPT-6 LunaHigh66.7<$0.01
  4. Muse Spark 1.3High33.4$0.10
  5. GPT-6.1 SolHigh33.3$0.07
  6. DeepSeek V4.1 FlashHigh33.3$0.03
  7. GLM-5.3 FlashHigh33.3<$0.01

Quality is the mean oracle score of the moves played, × 100. The best move scores 100; an illegal move or no move scores 0. Protocol boardbench-0.2.

Quality vs cost

7 of 7 models
  • Anthropic
  • OpenAI
  • DeepSeek
  • Z.ai
  • Meta
4050607080$0.001$0.01$0.1Claude Haiku 5.5 High · 66.7% · <$0.01 · Pareto frontierClaude Opus 5.5 High · 67.8% · $0.03 · Pareto frontierDeepSeek V4.1 Flash High · 33.3% · $0.03GLM-5.3 Flash High · 33.3% · <$0.01GPT-6 Luna High · 66.7% · <$0.01 · Pareto frontierGPT-6.1 Sol High · 33.3% · $0.07Muse Spark 1.3 High · 33.4% · $0.10Claude Opus 5.5 HighClaude Haiku 5.5 HighGPT-6 Luna HighMuse Spark 1.3 HighGPT-6.1 Sol HighGLM-5.3 Flash HighDeepSeek V4.1 Flash High
  • Pareto frontier: nothing plotted is both cheaper and better
  • Anthropic
  • OpenAI
  • DeepSeek
  • Z.ai
  • Meta

Horizontal axis is logarithmic.

ModelQualityLegalityCostn
Opus 5.5High67.8100.0$0.033
Haiku 5.5High66.7100.0<$0.013
LunaHigh66.7100.0<$0.013
Muse Spark 1.3High33.4100.0$0.103
SolHigh33.3100.0$0.073
DeepSeek V4.1 FlashHigh33.366.7$0.033
GLM-5.3 FlashHigh33.366.7<$0.013

Headline is the mean move score × 100. n is answers: puzzles × samples.

Look