The benchmark where AIs have to lie to win
Dwarkesh Patel ran about 386 games of the social deduction game Blood on the Clocktower, with AI models playing good and evil teams against each other, to measure how well they scheme, bluff, and detect deception. The piece walks through the rankings and some striking in-game plays.
- GPT-6 Astra is currently the strongest model in the benchmark, about 430 Elo above Claude Opus 4.6 (released February), winning roughly 92% of head-to-head games.
- OpenAI's newer models gained about 580 Elo over six months, mostly by playing the good side better: they weigh evidence holistically, while Anthropic's models tend to take reported information at face value.
- Some evil plays were sophisticated, such as a minion volunteering for execution and a backup Demon who exploited a trap set by the dying real Demon, showing long-horizon coordination and deception.
Dwarkesh Patel built a benchmark out of Blood on the Clocktower, a social deduction game in the Werewolf/Mafia family where one evil team, led by a Demon, tries to kill off the good team while good players try to execute the Demon during daily votes. Patel says he spent about $20,000 in API credits running 386 games in which one model played the whole good team and another played the whole evil team. His pitch is that the game tests things other benchmarks barely touch: bluffing that stays consistent over days of questioning, coordinating secretly with teammates, reasoning collectively when some participants are actively lying, and sometimes sacrificing yourself for the team.
The headline result is how fast things moved. Claude Opus 4.6, released in February 2026, was the first flagship model of that year in his rankings. GPT-6 Astra, released seven months later in September, is currently the strongest model in the benchmark, sitting about 430 Elo points above Opus 4.6. By Patel's arithmetic that works out to Astra winning roughly 92% of head-to-head games. He reads this as real progress in long-horizon coherence, multi-agent coordination, and strategic reasoning, not just a few points of benchmark noise.
What makes the piece worth reading is the play-by-play. In one game, GPT-6 Astra as evil had its Demon claim to be the Investigator, then let an evil minion volunteer for execution. A good player, the Undertaker, later confirmed the executed player had been evil, which made the fake Investigator the most trusted person at the table, and evil went on to win. In another, Opus 5.5 as evil set up a trap: an evil heiress with a backup-Demon power joined the mob against the real Demon, then proposed that if the game didn't end, the village should execute the claimed Saint, a player whose execution would instantly hand evil the win. The Saint accepted the deal, the Demon was executed, the game continued because the heiress took over, and the village then voted out the real Saint. Patel's examples are the core of the argument, since they show the models doing things like planting a conditional trap in public and then exploiting the other side's trust in it.