GPT-6 Astra leads Patel’s deception-game benchmark, with gains in spotting lies

An AI contest built around secret alliances and convincing lies produced a revealing result: the strongest model’s advantage was especially pronounced when uncovering deception. Dwarkesh Patel published results on November 5 from 386 games of Blood on the Clocktower, placing GPT-6 Astra first and attributing much of OpenAI’s improvement to stronger reasoning on the good team.

Two teams, unequal information

Blood on the Clocktower divides players into good and evil teams, each with special roles. Evil players know their teammates, including their leader, the Demon. Good players do not know who belongs to which side. Each day, players vote to execute someone; each night, the Demon secretly kills a player.

The good team must execute the Demon before evil kills off its opponents. That creates two different reasoning problems. Evil must coordinate believable stories and adapt them as new information appears. Good must combine clues while allowing for both deliberate lies and game mechanics that can make otherwise useful evidence misleading.

For the main rankings, each team consisted entirely of copies of one model, facing copies of another. Patel says he spent about $20,000 in API credits building the software and running games. He also replaced the human Storyteller—the person administering the game—with a programmatic system using manually set probabilities, aiming to keep discretionary rulings from influencing results.

A sacrifice builds trust; a failed attack misleads

One match illustrates the coordination the test can expose. Playing against Claude Opus 5.5, Astra’s Demon pretended to be an Investigator, a good character who receives clues about evil players. An evil teammate volunteered to be exposed because his power could still help after his death.

The good team executed that teammate. Another good player then verified that the dead player really was evil, making the Demon’s false Investigator claim look credible. The sacrifice helped turn the enemy leader into a trusted participant, protecting it while it continued killing good players. Evil won.

But Patel’s explanation for OpenAI’s overall advantage focuses more on skepticism than bluffing. He describes a paired example using Opus 5.5 and GPT-6.1, with the same roles assigned and the models switching sides. In both games, a good player called the Slayer attacked the actual Demon. The attack failed because evil had disabled the Slayer’s ability.

The Claude players treated the failed attack as clearing the suspect. The GPT players kept open the possibility that the attack had been sabotaged. By the following day, they had reconstructed what happened, executed the Demon and won. The distinction was not access to a different clue, but whether that clue settled the question.

A large lead, with narrower answers elsewhere

In Patel’s rankings, Astra finished about 430 Elo rating points above Claude Opus 4.6, a gap he translates into an estimated 92% win rate against that model. Astra’s lead over GPT-5.4 was larger: 580 points, corresponding to about 97%. Those are comparisons within this game benchmark, not measurements of general-purpose reliability.

Two other experiments produced less decisive answers:

• Mixed teams: In about 130 games, a team split between Opus 5.5 and GPT-5.6 Sol performed worse than teams of either model alone. The difference was not statistically significant.
• Stronger opponents: Evil’s win rate appeared to fall slightly as both sides became stronger, but Patel says the change was too small to distinguish from chance.

Patel has made his code available for others to investigate further questions. He argues that the competitive setup could resist benchmark saturation: rather than chasing a fixed score ceiling, models must keep overcoming increasingly capable opponents.