The author built an AI multi-agent benchmark based on the social deception game Blood on the Clocktower, enabling different models to play as the good and evil sides against one another and making the environment code publicly available. A ranking of 386 same-model matchups showed that GPT-6 Astra outperformed Claude Opus 4.6 by roughly 430 Elo, with a match win rate of about 92%. From GPT-5.4 in March to GPT-6 Astra, OpenAI gained 580 Elo.
Source:Dwarkesh Podcast · Read original
Leave a Reply