newsfilter
← Interviews
Dwarkesh Patel18 min

Which AI is the best liar?

Webinar, Other
  • The Blood on the Clock Tower benchmark provides unique visibility into creative bluffing, collective reasoning, and long-horizon strategizing, addressing current AI limitations in these specific areas.
  • GPT-6 Astra currently holds a 430 ELO point advantage over Claude Opus 4.6, winning approximately 92% of head-to-head games, and is expected to maintain this status until a superior model emerges.
  • A projected 580 ELO point gain over six months indicates OpenAI's newest models will win 97% of games against Claude Opus 5.5, driven by significant improvements in long-term coherence, multi-agent coordination, and strategic reasoning.
  • Future OpenAI models are expected to exhibit heightened skepticism toward individual evidence, adopting a holistic analysis of information over several days, whereas Anthropic models may process reported information more literally.
  • The evaluation benchmark is projected to avoid saturation as the performance bar is continuously reset by emerging smart models, though a hybrid team of different models might see a score rise due to unproven idiosyncratic biases.
  • As capabilities improve, the evil team's win rate is expected to decrease slightly, though this drift is currently too small to be statistically distinguishable from chance.
  • Emerging models may develop the ability to identify teammates by analyzing writing styles, a strategy that has not yet been observed in current iterations.
  • Updates to the benchmark will be published on clocktowerbench.com as new models are released, with future integration allowing users to play directly against AIs to test human win rates against top-tier models.