MafiaBench - Win Rate as Town: leaderboard
MafiaBench percentage of games won when the model plays all six townspeople of an eight-player Mafia game against two mafia played by another model, over a 15-round Swiss tournament (April 2025). Higher is better, 0-100 (%). Opponents differ by model because later pairings follow Elo.
Metric: Win rate as town (%). Source: www.mafiabench.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro | 52 | #454 |
| 2 | GPT-4o Mini | 51 | #598 |
| 3 | Claude 3.7 Sonnet | 51 | #259 |
| 4 | Claude 3.5 Sonnet | 49 | #362 |
| 5 | Gemini 2.0 Flash (001) | 36 | #365 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mafiabench-win-rate-as-town · How It Works · Data refreshed daily, snapshot 2026-11-05.