Dwarkesh Patel ran 386 games of Blood on the Clocktower with AI models. Models improved rapidly over months, with GPT-6 Astra outperforming Claude Opus 4.6 by 430 Elo. Astra wins 92% of games against Opus 4.6. OpenAI models gained 580 Elo points in six months. Astra excels at playing on the good side due to skepticism and holistic reasoning. Evil teams sometimes win through coordinated deception and strategic sacrifices. Hybrid teams of different models showed no clear performance benefit. The benchmark highlights AI strengths in long-term coherence and multi-agent coordination. It also reveals limitations in long-horizon strategizing. The test cost $20,000 in API credits.
What it unlocks
Measuring AI progress in long-horizon strategizing, multi-agent coordination, and strategic reasoning through a social deception game.
- 386 games played
- 430 Elo gap between GPT-6 Astra and Claude Opus 4.6
- 92% win rate for Astra vs Opus 4.6
- $20,000 spent on API credits
Sources