$20K of API credits: Blood on the Clocktower benchmark ranks AI liars by Elo
moultano · x · 2026-11-06
Dwarkesh Patel spent $20,000 in API credits having AI models play Blood on the Clocktower against each other, launching Clocktower Bench with 538 replayable games and full per-seat reasoning traces.
Key findings:
- Models improved dramatically this year at on-the-spot scheming, bluffing, and team-based detective reasoning;
- Ratings are Elo-based, with every model playing both the town (detection) and the Demon's team (deception);
- The two skills track closely: OpenAI's lead comes mostly from detection — its models improved as detectives about 4x faster than Anthropic's over the year, while deception skills are roughly level;
- The author notes OpenAI's newer models are clearly better than Anthropic's at this game.
The retweeter adds a spicy hypothesis: all-same-model teams might recognize and over-trust their own kind.
More from Models
- Yacine publicly asks insiders how many GPUs pretrained Inkling Small 2 — yacineMTB · 2026-11-06
- Autonomous AI R&D Runs 'Extremely Dumb Experiments', Researcher Reports — panickssery · 2026-11-06
- Claude Opus 5 pulls off multi-level roleplay: staying in character while knowing it's acting — repligate · 2026-11-06
- Inkling 2 Small hits 46.8% Pass@1 on APEX-Agents, up 13 points over predecessor — ziqiao_ma · 2026-11-06
- Dev complains Anthropic free credits can't be used for Code or terminal — RileyRalmuto · 2026-11-06
- Blind listening test of 62 raters ranks Gemini first, Llama last among 8 LLM composers — antiochIst · 2026-11-06