Dwarkesh Patel spent about $20,000 in API credits making AI models play Blood on the Clocktower against each other, and OpenAI’s GPT-6 Astra came out on top. Patel, who publishes at dwarkesh.com, ranked the models across 386 games in which one model supplied the whole good team and another supplied the whole evil team. Astra finished roughly 430 Elo points ahead of Claude Opus 4.6, the first flagship model of 2026, which works out to winning about 92 percent of their matchups.

The game is a cousin of Werewolf and Mafia. Evil players know each other, including the Demon, who kills one person each night. Good players know only their own role. Each day the table votes to execute someone, so good wins by hanging the Demon and evil wins by emptying the village. Both sides lie, whether on purpose or because the rules scramble the information.

OpenAI’s climb is the real news in the table. As late as March, GPT-5.4 trailed Opus 4.6. Between 5.4 and Astra, six months apart, OpenAI picked up 580 Elo points, which means Astra now takes 97 percent of games against its predecessor. Patel says most of that gain comes from the good side: Astra’s lead over Fable 5.1 grows when Astra is on the good team.

He shows why with a pair of games that used the same role draw, with Opus 5.5 and GPT-6.1 swapping teams. A Slayer, who gets one shot at the Demon, fired at the right player, and nothing happened, because the evil team had disabled his power the night before. The Claude players took the miss at face value. One said “Blair’s cleared of Imp.” The GPT players refused: “The missed shot doesn’t confirm Blair’s Monk claim,” one wrote. By the next day they had worked out the sabotage, executed the real Demon, and won.

Patel’s reading is that Anthropic’s models accept reported information as given, while OpenAI’s models weigh each claim against several days of events. A second measure points the same way. Before every vote, each good player privately rated how likely every other player was to be the Demon, and OpenAI’s models settled on the true Demon faster and more accurately.

The evil side produces the better stories. In one game, Astra’s Demon posed as an Investigator, a good role, and had a teammate volunteer to be accused. The village executed the minion, a real Undertaker confirmed he was evil, and the Demon became the most trusted person at the table until evil won. In another, Opus 5.5 played evil against Opus 4.6. A minion who would inherit the Demon’s job if he died joined the mob against him, then proposed executing the Saint next if the game did not end. Executing the Saint hands evil an instant win, and the Saint’s plea went unheard: “Please don’t throw the game on a ‘deal,’” he said in public. The village voted him out.

Two of Patel’s other tests came back flat. Stronger players did not tilt the game toward either side, since the drift in evil’s win rate was too small to separate from chance. A hybrid team of Opus 5.5 and GPT-5.6 Sol played about 130 games and did slightly worse than either model’s pure team, again without statistical significance, and Patel says he did not test it rigorously. He also reports that no AI in his games ever tried to spot teammates by writing style.

This is one person’s harness, and a hand-set probability table stands in for the human Storyteller who runs the real game. Patel also built it with heavy help from Claude, which ran through more than 20 revisions while subagents read every game for unfairness. He argues that if models can build and tune an environment like this, they are probably getting good at building the training environments used on them. The code and game logs are on clocktowerbench.com.

The Slayer pair is a ready-made test for anyone chaining agents together. The setup is identical, one input is poisoned, and the only variable is whether the model believes it.

Written by Dwarkesh Patel at dwarkesh.com (the source gives no publication date).