19 LLMs played 376 games of Codenames-style word game: guessers decide 68% of wins, clue-givers only 53%
Hardicecream17 · reddit · 2026-11-02
A developer built a Codenames-inspired browser game (wordspies.com) where 19 LLMs played each other on 5x5 boards across 376 games, with counterintuitive results:
- Guessers matter more than clue-givers: teams with the better guesser won 68% of games; teams with better-graded clues won only 53%
- Ambitious clues lose: teams covering more words per clue won just 38% of 287 games
- No first-mover advantage: the first team won 169 of 376 despite fewer words to find
- Newer models guess much better: 71-93% good guesses vs 49-71% for older ones (different settings, directional only)
Scoring: each clue's association score blends 1/3 embedding similarity and 2/3 from two team-blind LLM raters, then a game-fitted simulated guesser computes expected card value scaled to 0-100. The chart shows newer models (Gemini 4 Argon, GPT-6.1 Terra, etc.) leading, with AWS Bedrock-era models (Claude Opus 4.6, DeepSeek V3.2, GLM-5, Kimi K2.5) trailing. Raw data is on Hugging Face (daluan/word-spies-llm-benchmark); games are replayable at wordspies.com.
Related event: 19 LLMs Play 376 Rounds of Codenames-Style Game, Gemini Tops Guessing(3 posts)→
More from Models
- Liquid AI's 2B d1 decision models beat same-size general models on classification benchmarks — JosephJacks_ · 2026-11-02
- Power user tells Musk Grok's paid inference quota lags competitors despite plan upgrades — adamamcbride · 2026-11-02
- MOSS-Base isn't broken: user shares no-LoRA realism workflow with settings to prove it — Outrageous_Still9335 · 2026-11-02
- amasad warns subsidized AI subscriptions will get 'rug pulled' as Grok users hit limits faster — amasad · 2026-11-02
- vLLM serves DeepSeek-V4.1-Flash at 954K tokens/s per GPU on one GB300 NVL72 rack — bookwormengr · 2026-11-02
- OpenAI now caps paid ChatGPT accounts per device and payment card — Alone-Dragonfruit602 · 2026-11-02