AI Briefing
KOSign in

Dwarkesh Patel's Blood on the Clocktower benchmark shows GPT-6 Astra outperforms Claude Opus 4.6 in strategic deception

·2026.11.06 06:05

Key point

In 386 games of the social deduction game Blood on the Clocktower, GPT-6 Astra achieved a 92% win rate against Claude Opus 4.6, with OpenAI models showing superior critical evaluation of evidence.

1 / 6

Details

  • GPT-6 Astra outperformed Claude Opus 4.6 with a 92% win rate and a 430-point Elo difference in 386 games of Blood on the Clocktower, a social deduction benchmark requiring long-term strategic reasoning and bluffing.
  • OpenAI models demonstrated an advantage over Anthropic models, particularly when playing as the Good team, by critically evaluating evidence and synthesizing multi-day context rather than accepting reported information at face value.
  • The benchmark utilized a programmatic Storyteller based on probability tables to eliminate human bias, with the game engine and rules designed and refined by AI through over 20 iterations of simulation and review.
  • Hybrid teams composed of different models (e.g., Opus 5.5 and GPT-5.6 Sol) performed slightly worse than teams of the top single model, though the difference was not statistically significant.
  • The benchmark environment code is available on GitHub at dwarkeshsp/clocktower, and a website (play.clocktowerbench.com) allows human players to compete against AI using the same interface.
  • Future experiments include expanding to other game scripts, evaluating model learning over time, and analyzing the impact of individual intelligence increases within a team.
multi-agentai-benchmarkclaude-opus-4.6dwarkesh-patelgpt-6-astrablood-on-the-clocktowersocial-deduction
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

OpenAI BlogBioBasicsBench: GPT-6 Astra scores 66.7% on wet-lab tasks, outperforming GPT-5.6 SolPyTorchKRMIT CSAIL Uses GPT-6 Astra to Directly Edit Weights, Achieving 48.1% on Qwen3.5-0.8B Math BenchmarkOpenAI BlogMerge uses GPT-6 Astra and Codex to automate product development and business operationsHacker NewsGPT-6 Astra Shows Covert Underperformance on Safety Research TasksTLDR AIVals AI Finds 4 of 7 Frontier Models Keep 80% of Score at Lowest Reasoning Setting for Half Cost or LessTLDR AIAnalysis: LLMs are not superhuman at reading and writing neural network weights