Podcaster Dwarkesh Patel spent $20,000 on API calls to run a tournament of 11 models playing Blood on the Clocktower, a social deduction game where a minority of evil players know each other while the good majority must reconstruct the truth from nightly hints. To separate the two skills, one model played the entire good camp and another the entire evil camp, then they swapped roles, across 386 rounds. GPT-6 Astra took first place with an Elo of 1859 against 1549 for Claude Opus 5.5, and it also led on detecting lies (1940 vs 1500) and on producing them (1806 vs 1585).

Individual games showed recognisably human patterns: Astra sacrificed its own ally to the vote to earn trust, while Opus 5.5 took a rival’s claim at face value where Astra rechecked the evidence and found the bluff. Across this year’s releases, OpenAI is advancing faster at spotting disinformation while both labs improve at deception at a similar pace.

The harness is open at github.com/dwarkeshsp/clocktower, and anyone can play against the models at play.clocktowerbench.com.

Related: OpenAI Launches GPT-6 Astra with Strong Benchmark Results, Anthropic Releases Claude Opus 5.5

Blood on the Clocktower benchmark · Dwarkesh Patel’s tournament video