Someone spent and #036;20,000 AIs can lie to each other in Werewolf Podcaster Dwarkesh Patel paid about $20,000 in API credits to make the borderline models play Blood on the Clock - a social deduction game like a wolf - against each other for 386 games. One model plays all the good teams, another plays all the bad teams. OpenAI's GPT-6 leads the board with 1,859 hits against Cloud Opus 4.6, winning about 92% of the time. Dewarsh says the edge is mainly absorbing lies, and not telling better things. Full ratings and replays are at Watchtowerbench.com. You can also play AIs yourself. The terms: