Dwarkesh Patel's Clocktower Bench puts GPT-6 Astra on top of AI models playing a lying-and-detection game
Patel ran 386 games of Blood on the Clocktower between models, spending about $20,000 in API credits; he reports Astra wins about 92% of games against Claude Opus 4.6, a roughly 430 Elo gap, with OpenAI's lead mostly on the good side.
Why it matters
Blood on the Clocktower is a social-deduction game in which an evil team lies to a good team that must find its Demon, so it tests long-horizon coordination, bluffing and evidence weighing together. Patel replaced the human storyteller with a programmatic one and rated 11 models with a Bradley–Terry fit; he says OpenAI's models gained about 580 Elo between GPT-5.4 and GPT-6 Astra, mainly by treating single pieces of evidence more sceptically. A small mixed-team test of Opus 5.5 and GPT-5.6 Sol did better than neither side. Watch for new models added to clocktowerbench.com.
Key facts
- 386 games with one model as the whole good team and another as the whole evil team
- Astra wins about 92% of games against Claude Opus 4.6, roughly 430 Elo apart
- OpenAI models gained about 580 Elo between GPT-5.4 and GPT-6 Astra
- About $20,000 of API credits spent
- 11 models rated; 1500 Elo is their average
- Mixed Opus 5.5 / GPT-5.6 Sol teams played 132 extra games and rated a little below their members' average