Key Takeaways

  • Dwarkesh Patel spent $20,000 in API credits running automated matches of the social deception game Blood on the Clocktower between Claude Opus 5.5 and GPT-6.1.
  • In a mirrored matchup, Claude on the Good team cleared an evil Demon simply because a Slayer shot failed to kill him, taking the surface observation as absolute truth.
  • GPT-6.1 on the Good team deduced that an evil Poisoner had disabled the Slayer the night before, correctly rejecting the surface claim and executing the Demon to win.
  • Daily private probability tracking showed OpenAI models identified the real Demon faster and with higher confidence across rounds than Anthropic models.
  • Claude autonomously wrote and balanced the game's execution harness, even though its own game-playing logic fell for enemy deception.

Claude Believed the Surface Evidence

Social deduction games punish literal thinking. Patel discovered this when setting up a mirrored match where Claude Opus 5.5 and GPT-6.1 swapped roles to isolate how their reasoning diverged under identical rules.

The critical moment came down to an interaction between three roles: the Slayer, the Poisoner, and the Demon. The Slayer can make one public attempt to shoot a player; if that player is the Demon, the target dies immediately. In this match, the Slayer fired at an evil player named Blair. Blair did not die.

Claude, playing the Good team, immediately closed the case. If Blair did not die, Blair could not be the Demon. Patel noted this stark blind spot: “When the Claws were playing on the good side, they immediately accepted the conclusion that the previously suspected demon was innocent because after all, he didn't die.” (play from 8:57) Patel explained the underlying flaw in Anthropic's output: “Whereas enthropics models tend to take reported information at face value.” (play from 8:05)

Because Claude assumed observations reflect reality directly, it dismissed all suspicion against Blair. Blair claimed to be a Monk, and Claude bought the claim.

OpenAI Modeled Hidden States

When GPT-6.1 played the same Good team side, it approached the failed shot as one branch in an uncertain tree rather than settled proof. Patel quoted the model's inner deduction: “The missed shot doesn't confirm Blair's monk claim. In that specific world, this myth rules out Blair as imp. Not in every world.” (play from 9:15)

The model recognized that a negative result could mean two different things. Either Blair was not the Demon, or an unseen ability had interfered with the Slayer. Patel explained the breakthrough: “By the next day, the GBTs had worked out what happened, which was that the evil team had sabotaged the Slayer on purpose.” (play from 9:25) An evil Poisoner had acted the night before, neutralizing the Slayer before the trigger was pulled. GPT-6.1 voted to execute Blair, securing the victory.

This gap was not an isolated lucky guess. Patel tracked quantitative beliefs throughout the matches. “Each day before the vote, every good player was privately asked how likely they thought it was that each of the other players was the demon. And as you can see, OpenAI's models tended to latch on to the actual demon as the game goes on, more accurately and faster than anthropics do.” (play from 9:54)

Anthropic's models excelled at structuring code. Claude wrote and balanced the test setup without human intervention. Yet when operating inside an environment containing active deception, it lacked the adversarial skepticism required to question its own inputs.

What to Do With This

Audit your multi-agent architecture this week for single-point truth assumptions. Inject one corrupted or adversarial input into your agent pipeline and check whether downstream models accept the output as ground truth or verify alternative explanations before acting.