Scrabble

A board, the mover’s rack and the unseen tiles, ENABLE lexicon. Simulation scores every legal play by equity, not just points; the answer is a square and a word. See Methodology.

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

6 of 6 models
  • Anthropic
  • OpenAI
  • DeepSeek
  • Z.ai
203040506070$0.001$0.01$0.1$1Claude Haiku 5.5 High · 32.2% · $0.01Claude Opus 5.5 High · 14.5% · $0.23DeepSeek V4.1 Flash High · 22.1% · $0.06GLM-5.3 Flash High · 56.0% · $0.02 · Pareto frontierGPT-6 Luna High · 36.1% · $0.01 · Pareto frontierGPT-6.1 Sol High · 53.9% · $0.15GLM-5.3 Flash HighGPT-6.1 Sol HighGPT-6 Luna HighClaude Haiku 5.5 HighDeepSeek V4.1 Flash HighClaude Opus 5.5 High
  • Pareto frontier: nothing plotted is both cheaper and better
  • Anthropic
  • OpenAI
  • DeepSeek
  • Z.ai

Horizontal axis is logarithmic.

Table

ModelQualityLegalityCostn
GLM-5.3 FlashHigh56.0100.0$0.023
SolHigh53.966.7$0.153
LunaHigh36.166.7$0.013
Haiku 5.5High32.266.7$0.013
DeepSeek V4.1 FlashHigh22.133.3$0.063
Opus 5.5High14.533.3$0.233
Muse Spark 1.3High0.00.0—3

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

Look