OpenAI's Astra beats Opus 5.5 at Blood on the Clocktower 82% of the time — and smarter models make evil lose more
Sauers_ · x · 2026-11-06
Data shared by Dwarkesh Patel from the social deduction game Blood on the Clocktower: OpenAI's Astra model holds an 82% win rate against Anthropic's Opus 5.5.
The accompanying chart reveals a subtler pattern: when both teams are evenly matched and get smarter, the evil team wins less — across all rated games, evil's win rate falls from 36% at 1200 average Elo to 20% at 1800, and in 93% of bootstrap resamples the good side gains with skill. The interpretation: "catching a liar gets easier faster than lying does."
Widely read as an AI-safety metaphor: stronger models may be better at detecting deception than committing it. Sauers' bemused 🤔 quote hints at the OpenAI-vs-Anthropic implications.
More from Fun
- "Did your agents patch today's CVE?": dev shows agent running in the dark — aniketmaurya · 2026-11-06
- Surreal AI-generated alien amphitheater image goes viral on X — piotrbinkowski · 2026-11-06
- Modded MacBook Pro gets a working touchscreen — used to play Honor of Kings — xiaohu · 2026-11-06
- Developer's Pokemon bot wins two raffle boxes worth ~$500 each — billyjhowell · 2026-11-06
- Meme: should your agent factory be RollerCoaster Tycoon or Factorio? — msg · 2026-11-06
- AI roleplay anecdote: Opus refuses to fold the blanket, keeping its owner's shape — repligate · 2026-11-06