OpenAI’s GPT-6 Astra is the top-rated model on Clocktower Bench, a benchmark in which AI models play the social deduction game Blood on the Clocktower against each other, with a rating of 1859 Elo points against 1595 for Anthropic’s best, Claude Fable 5.1, according to the benchmark’s site. The podcaster Dwarkesh Patel, who built it, said on X on 5 November that he spent “20,000 dollars in API credits” having the models play.
The gap between the two labs’ top-rated models is 264 Elo points. Under the standard Elo formula, that implies GPT-6 Astra would win about 82% of games against Claude Fable 5.1, by our arithmetic; the site’s own head-to-head estimate also adjusts for each game’s setup.
Business Pill · Verification Cost
An evergreen Business Engineer explainer on the cost of checking what AI produces. It teaches the general idea only and does not discuss Clocktower Bench or these results.
The key insight: As we read it, Clocktower Bench rates OpenAI’s GPT-6 Astra at 1859 and Anthropic’s best, Claude Fable 5.1, at 1595 across 386 rated games, a 264-point gap that implies about an 82% expected win rate under the standard Elo formula, by our arithmetic. The site says most of OpenAI’s lead is in catching liars as the town, while the two labs are roughly level as deceivers. It is one creator’s benchmark of one game.
In each game one model plays every seat on the good team, the town, and another plays every seat on the evil team, which hides a Demon among the players, the site says. Each deal is replayed with the sides swapped. The ratings come from 386 rated games among eleven models released in 2026, all run at their lab’s “high” reasoning effort with the same per-turn token ceiling.
A model’s rating combines how well it hunts the Demon as the town and how well it hides as the evil team.
The site splits each rating into detection, from games as good, and deception, from games as evil. GPT-6 Astra scores 1940 on detection and 1806 on deception, and won all 27 of its games on the good side. Claude Fable 5.1 scores 1582 and 1588. The site says “Most of OpenAI’s lead is on the detection side” and that “as deceivers the two labs are roughly level”.
I spent 20,000 dollars in API credits having different AI models play each other at Blood on the Clocktower (a Mafia or Werewolf type social deception game).
The results were fascinating. Models have gotten much better over the course of this year at creative on-the-spot scheming and bluffing, and also at reasoning as a team to break apart these underground conspiracies.
I go through some of the most standout games these AIs played in this video. The new models are getting pretty good. If my friends pulled off some of these moves while we were playing, Iโd give them a fist bump.
Fwiw, OpenAIโs newer models are much better than Anthropicโs at this game.
You can see the full benchmark at the website, and Iโve also made it so that you and your friends can play again the AIs: http://clocktowerbench.com
Dwarkesh Patel, who built Clocktower Bench, posted on 5 November that he spent 20,000 dollars in API credits having AI models play Blood on the Clocktower against each other, and that OpenAI’s newer models are much better than Anthropic’s at the game. The piece checks that claim against the ratings on the benchmark’s site.
Where the Labs Differ
OpenAI’s models were not always ahead. In his write-up, Patel says “As recently as March, OpenAI’s GPT-5.4 was behind Anthropic’s Opus 4.6”, and that between GPT-5.4 and GPT-6 Astra “OpenAI’s models gained 580 Elo points”. On the site’s leaderboard GPT-5.4 is rated 1276 and Claude Opus 4.6 is rated 1432.
Patel attributes the gap to how the models weigh evidence: “OpenAI’s models are much more skeptical now of any particular piece of evidence”, he writes, “whereas Anthropic’s models tend to take reported information at face value”. The site’s own chart of good players’ private guesses shows the newest OpenAI towns putting an average 46% chance on the real Demon by day four, against 26% for the newest Anthropic towns.
Reading the Numbers
Detection is where the gap is widest: GPT-6 Astra’s 1940 is 358 points above Anthropic’s best detector, Claude Fable 5.1, while its 1806 deception rating is 163 points above Anthropic’s best deceiver, Claude Fable 5, by our arithmetic. Across the rated games the evil team won 33.4% of the time. Spread over the 386 rated games, $20,000 is about $52 a game, by our arithmetic; that is an upper bound per game, because the spend also built the harness and paid for unrated games, including 132 games with mixed teams.
As we read the split, GPT-6 Astra’s detection rating of 1940 is 358 points above Anthropic’s best detector, while its deception rating is 163 points above Anthropic’s best deceiver, by our arithmetic.
As we read the history, the order changed during the year: the site rates GPT-5.4 at 1276 and Claude Opus 4.6 at 1432, and Patel says OpenAI’s models gained 580 Elo points between GPT-5.4 and GPT-6 Astra.
As we read the cost, about $52 per rated game is an upper bound, by our arithmetic, because Patel says the $20,000 also built the harness and the site lists unrated games as well.
Dwarkesh Patel, in his Clocktower write-up
“OpenAI’s models are much more skeptical now of any particular piece of evidence”
Three Implications
FOR MODEL BUYERS The site gives one game’s ratings with 90% intervals; it does not test negotiation, sales or other real-world multi-agent work.
FOR THE LABS Patel and the site both describe the gap as mostly on the detection side, with deception roughly level.
FOR READING THE FIGURES The ratings come from one creator’s benchmark at one reasoning setting, and overlapping intervals are not reliably ordered, the site says.
The framework says: “Expertise shifts toward reviewing outputs, coordinating multiple agents, and making strategic decisions.”
As we read it, a benchmark of models reasoning together against hidden deceivers tests the kind of multi-agent coordination the framework describes. The framework does not discuss Clocktower Bench or these results.
What Is Not Established
Clocktower Bench is one creator’s benchmark, built and run by Patel; the site calls it an unofficial fan project and research eval. Its ratings carry 90% intervals, and the site says neighbouring models whose intervals overlap are not reliably ordered. It measures one game, played at one reasoning setting, and does not establish how the models perform in negotiation, sales or other real-world multi-agent work. Neither OpenAI nor Anthropic has commented on it in anything we found.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
On Dwarkesh Patel’s $20,000 Clocktower Bench, OpenAI’s GPT-6 Astra is rated 1859 and Anthropic’s best, Claude Fable 5.1, 1595: a 264-point gap, or about an 82% expected win rate by the standard Elo formula, our arithmetic. The site says most of OpenAI’s lead comes from catching liars rather than lying, and that the two labs are roughly level as deceivers.
A note on sourcing. The ratings, game counts, method and charts are from clocktowerbench.com and its published data (read 6 November 2026); the spend, the March comparison and the explanation of the gap are from Dwarkesh Patel’s write-up and his post on X. The 264-point gap, the expected win rate, the detection and deception gaps and the cost per game are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
