What happened
In a November 5, 2026 essay, Dwarkesh Patel describes running AI models through Blood on the Clocktower, a social-deception game related to Werewolf and Mafia. Players are assigned to good or evil teams and receive roles with special powers. Evil players know their teammates, including the Demon; good players begin without knowing who is on which side. Players discuss evidence, vote to execute someone during the day, and the Demon secretly kills at night. Evil aims to eliminate the good team; good aims to execute the Demon. The game therefore creates repeated decisions with private information, public claims, strategic deception and consequences that unfold over several rounds.
Patel says he spent about $20,000 in API credits building the benchmark harness and running the models. The central comparison used 386 games in which each team consisted of copies of one model. He reports that GPT-6 Astra, released in September, was the strongest model tested so far, around 430 Elo points above Claude Opus 4.6, which he identifies as the first flagship frontier model released in 2026, in February. Patel translates that gap into Astra winning about 92% of games against Opus 4.6. He also reports a 580-point Elo gain between OpenAI’s GPT-5.4 and GPT-6 Astra over six months, corresponding to Astra winning 97% of games against GPT-5.4.
The essay attributes much of the difference between newer OpenAI models and Anthropic models to play on the good team. In one comparison, models had to interpret a failed attempt by a Slayer role to kill a suspected Demon. Claude players treated the failure as clearing the suspect; GPT players considered whether the evil team had disabled the Slayer’s ability and continued to investigate. Patel says the GPT team eventually identified the sabotage and executed the Demon. He also describes a separate analysis in which good players privately estimated who was the Demon: OpenAI models identified the actual Demon more accurately and earlier as the game progressed.
To avoid human Storyteller discretion shaping benchmark outcomes, Patel replaced that role with a programmatic system: a probability table intended to let rule-allowed outcomes occur at odds chosen to resemble an experienced human Storyteller. He says AI systems helped build and revise the environment through more than 20 iterations, running games and having parallel subagents inspect whether the results seemed unrealistic or unfair. Patel invites others to use his code, browse game logs and play models through clocktowerbench.com. The benchmark uses one game script, Trouble Brewing. A small experiment with mixed teams of Claude Opus 5.5 and GPT-5.6 Sol—about 130 games—did not show a statistically significant advantage over teams made up of one model.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This does not yet fit a pattern New Runtime tracks. The essay explores multi-agent reasoning through a custom game benchmark, but it does not establish a durable shift in agent architecture, deployment or evaluation practice.
Why now
Patel frames the benchmark as a way to observe capabilities that are difficult to see in conventional tests, including long-term coherence, coordination and strategic reasoning. He also argues that AI-assisted construction of the game environment itself suggests models have improved at building and revising simulated environments. These are the essay’s interpretations of the results, not evidence that this game predicts performance in production systems.
Why it matters
For teams building agents, the useful contribution is less a leaderboard than a test design: an environment can make coordination, memory across rounds, belief revision and adversarial information visible in concrete traces. A team could inspect whether an agent distinguishes a failed action from proof that its target is innocent, or whether collaborators preserve a consistent strategy while new evidence arrives. Such traces may help diagnose behaviors that a single-turn benchmark cannot expose.
The caution is equally practical. Blood on the Clocktower rewards game-specific tactics, and the benchmark’s rules, prompt design and model selection shape its outcomes. A large Elo gap within this environment does not show that Astra will outperform Opus by the same margin in software engineering, customer support or other work. Builders should treat the results as a hypothesis generator and test their own systems against tasks that resemble their actual operating conditions.
What happens next
- The benchmark will receive comparisons involving newly released models. (3 months, 70%). Watch for: updated leaderboards or game logs at clocktowerbench.com.
Limits and uncertainty
These results come from a bespoke harness and one game script, Trouble Brewing. Patel says the mixed-model team experiment was small and not statistically significant; it cannot settle whether model diversity helps. The reported Elo comparisons also depend on the chosen roles, game rules, probabilities and evaluation setup. The essay presents examples and aggregate results, but the benchmark does not independently establish that game performance predicts success in real-world multi-agent work. Patel notes that another Blood on the Clocktower benchmark had been published separately, with different experiments; the two environments are not a direct replication.
Skill to build now
Choose one coordination failure your team wants to detect—such as agents treating unverified claims as facts—and turn it into a short, repeatable adversarial simulation. Record each agent’s evidence, belief updates and final decision, then inspect the traces for whether the system corrected its view when information changed. Patel’s public game logs and playable environment can provide a concrete reference for designing such tests, but adapt the scenario to your own workflow rather than importing its leaderboard as a proxy for production quality.
Reader takeaway
A social-deception game makes multi-agent strengths and failure modes unusually legible, and Patel’s results suggest substantial differences between models in this particular environment. Use the benchmark as an invitation to build richer evaluations—not as a shortcut for deciding which model is best for your product.