19 LLMs play 376 games of Codenames-style word game; Gemini 3.8 Flash guesses 93% right
Hardicecream17 · reddit · 2026-10-31
A developer built a Codenames-like party game where LLMs play every seat, logging 376 games across 19 models. GPT-6 Astra guessed right 91% of the time with almost no bad picks; Gemini 3.8 Flash was the most accurate guesser (93%, 0 bad guesses), Gemini 4 Argon had the top clue grade (67), and Claude Opus 5.5 trailed at 79%.
Key findings:
- The guesser decides the game: better-guessing teams won 68% of games, better-clue teams only 53%.
- Ambitious clues lose: teams covering more words per clue won just 38% of 287 games.
- Going first doesn't help: first teams won only 169 of 376.
Clue grading mixes Cohere embeddings (1/3) with two blind LLM raters (2/3), then simulates the turn with a guessing model fitted on real games, scaled 0-100 by win-probability impact. Playable at wordspies.com; author seeks feedback on the grading method.
Related event: 19 LLMs Battle Through 376 Rounds of Codenames-Style Word Game(2 posts)→
More from Models
- OpenAI reportedly to train on users' private sessions with no opt-out — evilsocket · 2026-10-31
- Heavy users burn $700 of compute on a $20 ChatGPT plan: subscription pricing is a bet — faraz0x · 2026-10-31
- High schooler's "Image-Space Rule Discovery" paper accepted to ACCV 2026 — HirokatuKataoka · 2026-10-31
- Microsoft's MAI-DxO beats generalist doctors 83.2% vs 36.1% on complex diagnosis, Science study finds — emmanuelvivier · 2026-10-31
- LLMs are fluent in others' work but unreliable when pressed on details — AlexTensor · 2026-10-31
- 576GB 'go f*ck yourself Sam Altman' mirror spreads as OpenAI docs leak — retr0jirachi · 2026-10-31