Clocktower Bench: models play Blood on the Clocktower

Dwarkesh Patel released Clocktower Bench, a leaderboard where frontier models play the social-deduction game Blood on the Clocktower against each other. One model plays every good (town) seat; another plays every evil (Demon team) seat; each deal is then replayed with the sides swapped. Ratings combine how well a model hunts as town (detection) and how well it hides as evil (deception).

In the essay, Patel writes that he spent about $20,000 of API credits to build the harness and run the games. That figure is the author’s claim in the essay (and on X); it is not a field on the bench site.

Elo snapshot (ratings.json)

The rated set covers 386 games among eleven models (Elo mean 1500). Combined ratings and 90% intervals from the site JSON:

Model	Combined Elo	90% interval
GPT-6 Astra	1859	1779-1983
GPT-6.1 Sol	1734	1684-1815
Claude Fable 5.1	1595	1537-1668
GPT-5.6 Sol	1591	1543-1647
Claude Opus 5.5	1549	1489-1612

Neighbours whose intervals overlap are not reliably ordered. The site says most of OpenAI’s lead is on the detection side, while as deceivers the two labs are roughly level. Across the year, OpenAI models improved as detectives about four times faster than Anthropic’s, per the site’s own trend note.

How to read it

These are models versus models inside a structured party game, not a claim that the same systems can deceive humans in the open. Setup effects (which Minions are dealt, player count) tilt outcomes, and mixed-team games are rated separately. Use the JSON, not relayed screenshots, when you cite Elo.

Why this matters

If you care about multi-agent coordination, long-horizon bluffing and collective reasoning, this bench surfaces skills that exam suites rarely score, with a public replay library and a fit you can inspect. Astra’s lead and the detection-versus-deception split are useful colour for comparing OpenAI and Anthropic stacks without turning a party game into an alignment proof. Treat the $20k figure as the author’s cost note, and keep the conclusion inside the game: who catches liars, and who lies best, when every seat is another model.