realset/hanabi-agent-bench

Hanabi Agent Benchmark

21 LLMs play the cooperative card game Hanabi as four-player agent teams. 140 games, same seeds, per-move referee records, scores and reference cost.

English report · 中文报告 · Data (CSV) · GitHub · v0.1.6, 2026-10-09 · by flatkey

Why Hanabi

Most LLM benchmarks score single answers. Hanabi tests whether a model can keep acting as a teammate: each player sees everyone's cards except their own, hints are scarce, three misplays end the game, and the team shares one score out of 25. Doing well needs rule-following, reasoning about what teammates know, coordinating through limited hints, and staying consistent over 40–60 moves.

All four seats run the same model. A local referee engine validates every action and records what each seat could see. The unit compared is the whole system: model ID × access route × prompt and request settings.

Results

Rank	Model	Effort	Mean progress (/25)	Per seed (101/102/103/104)	Lost all lives
1	Claude Opus 5.5	Low	20.50	20 / 21 / 19 / 22	0/4
2	GPT-6 Astra	Low	18.75	18 / 20 / 19 / 18	0/4
3	Claude Fable 5.1	Low	18.00	18 / 17 / 21 / 16	0/4
3	Gemini 3.1 Pro Preview	Low	18.00	17 / 20 / 16 / 19	1/4
5	GPT-6 Sol	Low	16.25	17 / 14 / 17 / 17	0/4
6	Gemini 3.8 Flash	Low	16.00	17 / 15 / 17 / 15	0/4
6	Qwen 3.8 Max	Single	16.00	15 / 16 / 15 / 18	1/4
8	Seed 2.1 Lite	Single	15.50	14 / 17 / 16 / 15	0/4
9	DeepSeek V4 Pro	Single	15.25	17 / 14 / 14 / 16	1/4
10	DeepSeek Flash	Low	14.25	9 / 19 / 18 / 11	2/4

All 21 models: results/baseline_scores.csv.

What stands out

• More reasoning effort mostly helps, but not always. Switching from Low to High raised the mean for 13 of 14 models (40 of 56 seed pairs improved). GLM-5.3 gained +6.75 points; GPT-6 Astra lost 0.50.
• Being right when you play is not enough. Claude Haiku 5.5 is correct on 94.4% of the cards it plays, above Claude Opus 5.5's 92.1%, yet scores 12.75 vs. 20.50. It spends its turns on hints and discards and plays too rarely (54 play attempts vs. 89).
• Good games look different. Games scoring ≥20 averaged 0.75 misplays over 60.75 moves; games scoring ≤10 averaged 2.96 misplays over 38.50 moves. In good games, 84.6% of hinted cards were played successfully on the receiver's next turn; in bad games, 37.9%.
• Cost and score don't line up. Claude Opus 5.5 costs about $1.00 per game for 20.7 points; GPT-6 Luna about $0.01 per game for 6.3; HY3 and Kimi K3 score about the same (13.7 vs. 13.3) at $0.07 vs. $0.38 per game.

What's in this repo

Path	Contents
report/en/report.md	Full report in English
report/zh/report.md	Original Chinese report with all 10 figures
results/baseline_scores.csv	21 models × 4 seeds, mean progress, ending types
results/high_minus_low.csv	Per-seed High − Low differences for the 14 paired models
results/low_vs_high_process.csv	Misplays, play success, critical discards, hint follow-through: Low vs. High
results/action_counts.csv	Plays, discards and hint outcomes per model
results/reference_cost.csv	Reference cost per game and per point (games with complete token usage)
results/paired_cost_low_high.csv	Low vs. High cost on the same seeds
results/model_registry.csv	Models, vendors, access routes, effort settings
figures/	English figures (rebuild with python scripts/make_figures.py)

Setup in one paragraph

Base five-colour Hanabi, 4 seats, all seats call the same model; shared seeds 101–104; 21 baseline configurations (Low for the 14 models that offer Low/High, the only tier for the other 7) plus High for the 14 paired models: 35 configurations, 140 games, every one reaching a rule-based ending (143 attempts: 140 endings, 2 technical failures, 1 cancellation). Score = final board progress 0–25; a game lost to three misplays keeps the progress it made. Reference cost = reported tokens × list prices on 2026-10-09, not actual bills. Models were called through flatkey, OpenRouter and vendor-specific routes; exact routes are in results/model_registry.csv.

Limitations

• Four seeds per configuration. Enough to show tiers and obvious lows, not to order models with close means. No significance is claimed.
• "Low" and "High" are request settings, not equal compute across vendors, and it is not verified that every aggregator route passes them through.
• Same-model teams amplify one strategy's strengths and weaknesses; results don't transfer directly to mixed or human teams.
• Process metrics are descriptive. A play after a hint does not prove the hint caused it.
• Raw trajectories and the engine are not in this release. The per-move archives exist and are planned for a later release; until then, the CSVs here are the reproducible part.

Citation

@misc{hanabi_agent_bench_2026,
  title  = {Hanabi Agent Benchmark: Mainstream LLMs as Cooperative Agents},
  author = {{flatkey}},
  year   = {2026},
  note   = {v0.1.6, 2026-10-09},
  url    = {https://github.com/mguozhen/hanabi-agent-bench}
}

Report and data: CC BY 4.0. Hanabi is a game by Antoine Bauza; this project is not affiliated with its publishers or with any model vendor.