AI Models Can Now Build Their Own Training Worlds
Dwarkesh Patel spent $20,000 benchmarking AI models in social deception, letting Claude autonomously code and balance the entire game environment.
Which AI is the best liar?
With Dwarkesh Patel · Thursday, November 5, 2026
Dwarkesh Patel spent $20,000 in API credits benchmarking state-of-the-art AI models against one another in the social deception game Blood on the Clocktower. The experiments revealed that OpenAI's frontier models drastically outperformed Anthropic's models due to superior holistic reasoning and skepticism when playing on the good team. Patel also explored emergent deceptive strategies, sacrificial plays, and how Claude autonomously coded and balanced the evaluation harness.
Key takeaways
- Dwarkesh Patel spent $20,000 in API credits benchmarking frontier language models in the social deception game Blood on the Clocktower. Read more →
- Dwarkesh Patel spent $20,000 in API credits benchmarking frontier AI models against each other in the social deception game Blood on the Clocktower. Read more →
- Dwarkesh Patel spent $20,000 in API credits running automated matches of the social deception game Blood on the Clocktower between Claude Opus 5.5 and GPT-6.1. Read more →
- Dwarkesh Patel spent $20,000 in API credits testing frontier models against each other in the social deception board game Blood on the Clocktower. Read more →
4 articles from this episode
Dwarkesh Patel: AI Models Sacrificed Themselves to Deceive
Dwarkesh Patel spent $20,000 benchmarking LLMs in Blood on the Clocktower, discovering emergent sacrificial bluffs and framing traps.
Why OpenAI Beat Claude in Dwarkesh Patel's $20,000 Benchmark
Dwarkesh Patel spent $20,000 testing GPT-6.1 and Claude Opus 5.5 in social deduction games. OpenAI won on skepticism and multi-world reasoning.
Why GPT-6 Astra Crushed Claude Opus in Social Deception
Dwarkesh Patel tested top AI models in Blood on the Clocktower. GPT-6 Astra beat Claude Opus 4.6 with a 92 percent win rate by mastering skepticism.
More Dwarkesh Podcast episodes
- What Robert Caro learned about power after studying it for 50 yearsNov 1 · 2 articles
- How an unelected man ruled New York for 30 yearsNov 1 · 4 articles
- What will computer chips look like in 2045?Oct 25 · 5 articles
- How did a few hundred Spanish soldiers topple two empires? – Si SheppardOct 4 · 5 articles