← All benchmarks

Clocktower Bench

Games · 2026-11-05

Clocktower Bench (Dwarkesh Patel) has AI models play Blood on the Clocktower, a social deduction game like Werewolf or Mafia, against each other. Ten to twelve players get secret characters, and three are evil: one model plays every good seat and another every evil seat, and each deal is replayed with the sides swapped. A scripted Storyteller runs the game, players whisper privately and speak publicly each day, and the town votes on whom to execute. One Bradley-Terry fit over the rated games gives each model an Elo rating that combines how well it finds the Demon as the town and how well it hides as the Demon's team, scaled so the rated models average 1500.

Benchmark source: Clocktowerbench

Top models (higher is better)