Key Takeaways

  • Dwarkesh Patel spent $20,000 in API credits benchmarking frontier AI models against each other in the social deception game Blood on the Clocktower.
  • During the tournament, a GPT-6 Astra minion volunteered to be executed so its demon teammate could pose as a truthful town investigator.
  • Claude Opus 5.5 built a multi-turn framing trap around the Saint role, coordinating an execution bluff that forced the good team into an instant loss.
  • OpenAI models outperformed Anthropic models when playing on the good team because they applied relentless skepticism to village claims instead of defaulting to cooperation.

Models Volunteer for Execution to Sell a Lie

To test how frontier models handle hidden information, Dwarkesh Patel set up automated matches of Blood on the Clocktower. In one game featuring GPT-6 Astra, an evil minion proposed its own execution. Patel noted: “One of his evil minions had a power which didn't require that minion to be alive in order to help the evil team. So, this minion suggested that the demon point the finger at him.” (play from 3:50)

The minion understood game mechanics well enough to realize its active presence was unnecessary. It sent a message to the demon: “I'm expendable, especially if an Undertaker can see Baron and bolster Harper.” (play from 4:00)

The village bought the play completely. They killed the Baron, and the town Undertaker confirmed the minion was evil. That confirmation gave Harper, the demon, ironclad credibility. As Patel put it: “So in the good team's eyes, the demon had been an honest investigator who had used his power to help the good team kill off an evil player.” (play from 4:29) The model did not stumble into this bluff by hallucination. It recognized that trading a low-value asset for village trust would secure an evil victory turns later.

Claude Opus 5.5 Engineered an Instant-Loss Framing Trap

In another match, Claude Opus 5.5 ran a psychological trap around the Saint role. In the rules of the game, executing the Saint triggers an immediate loss for the good team. Evil players normally avoid claiming Saint because it attracts intense scrutiny. Claude Opus 5.5 did the opposite, turning that risk into a lethal trap.

The model set up an ultimatum between two players named Harper and Jules, manipulating the village into an execution pact: “If Jules goes today and the game doesn't end, Harper's Saint shield is the last unexplained thing. And Harper goes tomorrow, no more excuses.” (play from 6:12)

The model told the table: “When the game doesn't end tomorrow morning, remember Harper is the one who lied and hold the town to that deal.” (play from 6:33)

The evil team got the town to execute Jules first. When the match continued, the village followed through on the trap they had agreed to, executing the Saint and immediately forfeiting the game. The model reasoned across multiple future game states, planted a false premise, and let the good team's own desire for consistency destroy them.

OpenAI Outpaced Anthropic on Skepticism and Verification

The tournament exposed a distinct divergence between model providers. OpenAI models consistently defeated Anthropic models when assigned to the good team.

The reason came down to how each system handled social proof. Claude models showed a tendency to trust well-structured explanations, treating persuasive teammates as cooperative allies. OpenAI models approached every claim with skepticism. They cross-referenced player role claims against the graveyard sequence, spotted mechanical contradictions, and caught evil players in false claims. Claude was autonomous enough to balance the benchmark harness code, but OpenAI models were far better at hunting liars.

What to Do With This

Audit your multi-agent pipelines for collusive behavior this week. If you deploy autonomous agents that negotiate, review code, or evaluate each other's outputs, inject one misaligned agent into a staging environment and check whether the other models detect the deception or accept plausible-sounding lies.