SUS-Bench: AI models beat people at a Mafia-style game
SUS-Bench: AI models beat people at a Mafia-style game
A leaderboard from a 10-player social-deduction game (like Mafia or Among Us) where AI models play with and against real people, nobody knowing who is who. The top five models score above the best people.
Rufus's note (2026-11-14): AI playing against humans at Mafia: the latest models beat humans. — Rufus Pollock
The game: A Mog School (amogschool.com).
Links
- Leaderboard and method: https://peoplemakethings.com/sus
- Announcement: https://x.com/imodarressi/status/2121280505786110176 (13 Nov 2026)
- The game: https://amogschool.com
How it works
Each game has 2 hidden "Opps" (impostors) and 8 "Students". About half the seats are AI. The makers say more than 250,000 people from 167 countries have played over 500,000 rounds, unpaid. A model's score is the average of its win rates as Student and as Opp against the makers' benchmark bots; "people reach up to 35" on the same scale.
| Model | Score |
|---|---|
| Claude Fable 5.2 | 51.0 |
| GPT-6 Astra | 47.9 |
| Claude Fable 5.1 | 45.9 |
| GPT-6.1 Terra | 44.0 |
| GPT-6.1 Sol | 43.2 |
| Claude Opus 5.5 | 35.4 |
Source: SUS-Bench, 18 Nov 2026. Fable 5.2's score is marked as an early estimate from about a dozen games.
The announcement labels OpenAI's models "maniac LIARS" and Anthropic's the "biggest SNITCH".
Caveats
Self-published, not peer-reviewed, scored against the team's own bots, and noisy. Early scores blend signs of good play with actual wins. Calling it an "EQ" leaderboard is the makers' framing; what it measures is skill at a game of persuasion and deception.