AI Models Compete in 'Mafia': GPT-6 Astra Outperforms Claude Opus 5.5

Podcaster Dwarkesh Patel spent approximately $20,000 in API credits to run a tournament of 386 games of Blood on the Clocktower among 11 models from 2026, where GPT-6 Astra took first place with an overall Elo of 1859, while Claude Opus 5.5 placed fifth with an Elo of 1549.

Compiled by Sergey KostenchukPublished 2026-11-08Updated 2026-11-08

2026-11-08 Research OpenAI

🎭 AI Played 'Mafia': GPT-6 Astra Wins

Podcaster Dwarkesh Patel spent approximately $20,000 on API credits and conducted 386 games of Blood on the Clocktower—a complex version of 'Mafia' where a probability table replaces the host. Among 11 models from 2026, GPT-6 Astra (OpenAI) took first place with an Elo of 1859: 1940 in lie detection and 1806 in lie creation. Claude Opus 5.5 (Anthropic) placed fifth with an Elo of 1549.

🌍 The benchmark addresses a blind spot in standard tests: multi-agent coordination, multi-step planning, and resistance to disinformation. In the games shown, Astra sacrificed an ally-minion to execution to cover for the demon, while Opus 5.5 believed others' words.

👤 The answer cannot be memorized here: Astra wins against Opus 5.5 in approximately 82% of games. The leaderboard and methodology are open—figures can be verified at clocktowerbench.com.

Source 1: https://www.dwarkesh.com/p/clocktower Source 2: https://clocktowerbench.com/

Sources