BoardBench

Overall

Models play full games against classical engines. The bar is the mean ideal-decision rate across Yatzi, Scotland Yard, and Splendor.

  1. Claude Fable 5Medium93.1$1.66
  2. Claude Sonnet 5Medium88.9$0.30
  3. Grok 4.5Medium85.0$0.06
  4. DeepSeek V4 Flash 0731High84.5$0.02
  5. Grok 4.6Medium83.2$0.07
  6. Kimi K3Medium82.2$0.14
  7. GPT-5.6 LunaHigh81.5$0.03
  8. GPT-5.6 SolMedium79.3$0.33
  9. Claude Opus 5Medium77.7$1.66

Quality vs cost

11 of 11 models
  • Anthropic
  • OpenAI
  • SpaceXAI
  • DeepSeek
  • Moonshot AI
7580859095100$0.01$0.1$1$10Claude Fable 5 Medium · 93.1% · $1.66 · Pareto frontierClaude Opus 5 Medium · 77.7% · $1.66Claude Sonnet 5 Medium · 88.9% · $0.30 · Pareto frontierDeepSeek V4 Flash 0731 High · 84.5% · $0.02 · Pareto frontierGPT-5.6 Luna High · 81.5% · $0.03GPT-5.6 Sol Medium · 79.3% · $0.33Grok 4.5 Medium · 85.0% · $0.06 · Pareto frontierGrok 4.5 Low · 75.0% · $0.07Grok 4.6 High · 76.5% · $0.07Grok 4.6 Medium · 83.2% · $0.07Kimi K3 Medium · 82.2% · $0.14Claude Fable 5 MediumClaude Sonnet 5 MediumGrok 4.5 MediumDeepSeek V4 Flash 0731 HighGrok 4.6 MediumKimi K3 MediumGPT-5.6 Luna HighGPT-5.6 Sol MediumClaude Opus 5 MediumGrok 4.6 HighGrok 4.5 Low
  • Pareto frontier: nothing plotted is both cheaper and better
  • Same model at different efforts
  • Anthropic
  • OpenAI
  • SpaceXAI
  • DeepSeek
  • Moonshot AI

Horizontal axis is logarithmic.