Summary
Dwarkesh Patel pits frontier AI models against each other in Blood on the Clock Tower, a social deception game similar to Werewolf and Mafia, to measure capabilities that most benchmarks miss: secret coordination, sustained bluffing, and collective reasoning under deliberate misinformation. He spent roughly $20,000 in API credits building the harness and running 368 games in which an entire good team is copies of one model and the evil team is copies of another.
The rankings show steep progress during 2026. GPT-6 Astra leads, with a gap of about 430 ELO points over Claude Opus 4.6, which translates to roughly a 92% win rate for Astra. The video walks through concrete plays, including an evil minion volunteering to be sacrificed so the demon’s false claims look credible, and the good team’s attempts to unravel such schemes.
A key finding is that OpenAI’s newer models are especially strong on the good side. They treat individual pieces of evidence skeptically and weigh them against everything that has happened, while Anthropic’s models tend to accept reported information at face value. A matched-setup comparison, where a Slayer’s failed shot was caused by sabotage, shows the difference clearly. The video also tracks how quickly each model’s suspicion converges on the real demon, offering a vivid look at long-horizon strategy and multi-agent behavior.
📺 Source: Dwarkesh Patel · Published November 05, 2026
🏷️ Format: Benchmark Test
