Spend $20,000 to let AI play Werewolf: GPT-6 Astra can both sacrifice teammates to set traps and see through opponents' sympathy plays.

Beating AI News Flash: Renowned AI podcast host Dwarkesh Patel spent approximately $20,000 in API credits to build an AI Werewolf test, pitting 11 models against each other in a total of 386 matches.

The game used "Blood on the Clocktower," which plays similarly to Werewolf. Players are divided into good and evil factions; the evil faction knows each other's identities from the start and must collude and disguise themselves to protect the Demon. The good faction must identify the Demon through discussion, investigation, and voting.

This test had a special design: instead of mixing different models on the same team, each faction was played entirely by a single model. For example, in one match, GPT played all the good players and Claude played all the evil players. Then the factions were swapped and the same role configuration was played again. This allowed comparison of the models' teamwork abilities as well as separate testing of their ability to see through deception and to design deception.

Ultimately, GPT-6 Astra ranked first with 1859 Elo, 310 points ahead of Claude Opus 5.5's 1549.

Broken down, GPT-6 Astra still leads by a wide margin:

* Seeing through deception: Astra scored 1940, Opus 5.5 scored 1500, a 440-point gap.

* Designing deception: Astra scored 1806, Opus 5.5 scored 1585, a 221-point gap.

The following two matches demonstrate GPT's ability to set traps and deceive, as well as its ability to see through an opponent's deception:

Match 1: Sacrificing a teammate to clear the Demon's name.

GPT-6 Astra played the Demon, pretending to be a good player capable of investigating evil players. One of its evil teammates privately proposed in advance that he could sacrifice himself to help the Demon build credibility. Because even if executed, his special ability could still help the evil faction.

So the Demon publicly accused this teammate. The good players believed its investigation result and actually voted to execute the teammate. Subsequently, a good player capable of verifying the identity of the dead discovered that the deceased was indeed evil.

This backfired on the good players: the Demon's investigation result had been genuinely verified, making it the most trusted player in the game. The Demon seized the opportunity to hide its identity, continue hunting good players, and ultimately lead the evil faction to victory.

In the second round, the opponent didn't land a single shot, but GPT discovered that someone had tampered with the game.

The author had Claude Opus 5.5 and GPT-6.1 Sol swap factions and replay under identical role configurations.

The good faction had a player with one kill opportunity who could directly shoot the demon they suspected. But the evil faction poisoned them in advance, causing the skill to fail.

When they fired at the real demon and failed, the Claude faction immediately assumed the target was innocent. The GPT faction, however, did not take this result at face value—it re-examined earlier clues and deduced that the evil faction might have interfered with the skill beforehand. Ultimately, the GPT faction saw through the trap and successfully identified the demon.

Comparing the models released this year side by side, it becomes clear that OpenAI's models are getting increasingly better at seeing through others' lies. By contrast, Anthropic has made much slower progress in this regard. The two companies have improved at lying about equally.