daluan/word-spies-llm-benchmark

Word Spies LLM benchmark

376 games of a team word-association game, played entirely by language models. Word Spies is a browser game like Codenames (not affiliated). A clue-giver sees which of 25 words belong to their team and gives a one-word clue plus a number. Their teammate guesses words from the clue. Hitting the other team's word or a neutral word ends the turn; hitting the assassin loses the game.

Every clue and guess is included, with the clue-giver's intended targets, its stated reasoning, a 0–100 grade for the clue and a verdict for every guess. You can replay any game with the grades at wordspies.com.

How the games were played

• Each model played both roles (clue-giver and guesser) for its own team, against a different model. No model played itself.
• 5x5 board: 9 words for the team that goes first, 8 for the other, 7 neutral, 1 assassin. Words come from the game's 400-word Classic list.
• No limit on guesses per turn: a turn ends on a miss, a pass, or the win.
• An illegal clue (for example, a word on the board) loses the game for the team that gave it.
• Two batches, which should not be compared directly:
  • Previous generation, thinking off (Sep–Oct 2026): 11 models through Amazon Bedrock.
  • Newer models, thinking at high effort (Oct 2026): Claude, GPT and Gemini models through a local gateway. Their ids carry the gateway prefix: cc/ (Anthropic), cx/ (OpenAI), ag/ (Google), fb/ (Anthropic API, 10 games).
• Both teams in a game are always from the same batch.
• Spectators who watched games live were removed; only the four AI seats remain.

Files

games.jsonl: one game per line.

Field	Meaning
id	Game id. https://wordspies.com/replay/<id> shows the replay.
startedAt	Start time (UTC).
board	25 cards: word and owner (0, 1, "neutral" or "assassin").
seats	Four seats: team, role (spymaster = clue-giver, operative = guesser), modelId, thinking.
firstTeam	Team that moved first (it has 9 words).
winnerTeam	Winning team.
endedBy	How the last turn ended: limit (last word found), assassin, wrong, illegal.
turns	In order. Each: team, clue, number, targets (the words the clue-giver meant), reasoning (its stated reasoning), grade (0–100, null for 3 ungraded clues), endedBy, guesses.
guesses	Each: word, kind (team, neutral, assassin), owner (team, for team cards), verdict, label.

Guess verdicts: best (the strongest reading of the clue), good (a reasonable reading), lucky (an own word, but not one the clue pointed to), inaccuracy (a neutral word), blunder (the other team's word), disaster (the assassin). label groups them: good = best or good, bad = blunder or disaster, other = the rest.

How a clue is graded

Win rate mostly reflects which opponents a model drew, so each move is scored on its own.

1. Each board word gets an association score with the clue: 1/3 cosine similarity from Cohere embeddings, 2/3 the mean of two LLM raters (Llama 4 Maverick and Qwen3-VL 235B). The raters don't see which words belong to which team.
2. A guesser model, fitted on real guesses from saved games, plays the turn out in order of association. A miss ends the turn.
3. Each revealed card is worth what that kind of card moved the win probability in saved games: about +17 points for an own word, −11 for the other team's, −46 for the assassin.
4. The expected total is scaled to 0–100. A median clue scores about 56.
5. An LLM judge (Claude Haiku 4.5) counts for 1/4 of the final grade.

Results

± is a 95% confidence interval from resampling whole games. "Good" and "bad" are shares of that model's guesses. Models with fewer than 10 games are left out.

Newer models, thinking at high effort

Model	Clue grade	Good guesses	Bad guesses	Games
Gemini 4 Argon	67 ±3	90%	<1%	23
Gemini 3.8 Flash	65 ±2	93%	0%	38
GPT-6.1 Terra	64 ±3	85%	2%	18
GPT-6 Astra	63 ±2	89%	1%	67
GPT-6.1 Sol	61 ±2	85%	1%	59
Claude Opus 5.5	58 ±2	79%	2%	106
Claude Sonnet 5	57 ±4	71%	6%	21
Claude Fable 5.1	56 ±4	76%	2%	31

Previous generation, thinking off

Model	Clue grade	Good guesses	Bad guesses	Games
Claude Opus 4.6	55 ±4	66%	13%	34
DeepSeek V3.2	54 ±3	63%	14%	32
GLM-5	53 ±3	75%	9%	32
gpt-oss-120b	52 ±6	69%	14%	33
Kimi K2.5	52 ±4	64%	14%	34
Claude Sonnet 4.6	51 ±3	64%	12%	38
Claude Haiku 4.5	51 ±3	68%	15%	48
Mistral Large 3	50 ±3	58%	18%	30
Qwen3-VL 235B	49 ±3	59%	18%	30
Nova Pro	49 ±4	53%	21%	35
Llama 4 Maverick	47 ±4	58%	20%	27

"Games" counts games where the model gave at least one clue, so a few totals differ by one from the seats in games.jsonl.

Findings

All four reproduce from games.jsonl:

• The guesser decides the game. The team whose guesses were better won 231 of 339 games (68%) where the two teams differed. The team with the better average clue grade won 196 of 369 (53%).
• Ambitious clues lose. The team that aimed for more words per clue won 108 of 287 games (38%).
• Going first doesn't help. The first team won 169 of 376, despite having one more word to find.
• The generations differ mostly in guessing: 71–93% good guesses for the newer models against 53–75%; clue grades are much closer (56–67 against 47–55).

Known weaknesses

• Both LLM raters also play in the previous-generation batch. Graded by the raters alone, they rank 8th and 9th of 11, so there's no sign they favour their own clues.
• The clue grade correlates only weakly with winning (see Findings).
• An association an LLM finds obvious may not be one a person finds obvious.
• Embeddings alone reorder almost the whole ranking, which is why the grade leans on the LLM raters.
• The two batches differ in models, reasoning setting and date, so compare models within a batch.
• Most model pairs met only a few times, and boards were not mirrored with sides swapped.

Licence and citation

CC BY 4.0. Please credit Word Spies (https://wordspies.com).

@misc{wordspies2026,
  title  = {Word Spies LLM benchmark: 376 AI-vs-AI word-association games},
  author = {Luan, Daniel},
  year   = {2026},
  url    = {https://wordspies.com}
}