Which AI is the best liar?

*Blood on the Clocktower* serves as a sophisticated benchmark for evaluating AI capabilities in long-term strategic reasoning, creative bluffing, and multi-agent coordination. Recent testing reveals that models like GPT-6 Astra have significantly outperformed predecessors, achieving higher Elo ratings through improved coherence and skepticism toward deceptive information. Unlike traditional benchmarks, this environment forces agents to navigate complex social dynamics, such as identifying hidden conspiracies and adapting to intentional misinformation. The development of this programmatic game engine, largely assisted by AI, demonstrates that models can effectively iterate on their own testing environments. Future research opportunities include analyzing how models accumulate experience from previous matches, the impact of diverse model teams, and the performance gap between AI and human players in social deception scenarios.