Podcast host Dwarkesh Patel spent about $20,000 in API fees to have several cutting-edge models play Bloodstained Clocktower against each other for a total of 386 rounds.
"Bloody Clock Tower" is a social reasoning game of Werewolf genre. The rules are much more complicated than Werewolf, the information is more fragmented, and there is more room for lying and detecting lies. In the experiment, one model played the entire good camp, and the other model played the entire evil camp.
At the top of the list is OpenAI’s GPT-6 Astra, with Elo around 1859. Against Claude Opus 4.6, its win rate is around 92%.
Dwarkesh’s observation is interesting: Astra wins mainly by spotting lies, not by being better at lying. In other words, its advantage is on the reasoning side, finding contradictions in a bunch of mixed true and false statements.
The complete rankings and game replays are on clocktowerbench.com, and you can also go out and play a few games with these AIs yourself.
Opus 4.6 is no longer Anthropic’s latest model, and this comparison cannot directly tell which of the two current flagships is better. The model versions on the list vary, so it’s best to check the specific models when looking.
This type of benchmark score is difficult, but it is closer to a real-world ability than doing questions: making judgments when others intentionally mislead.