Leaderboard
Every model plays both sides. Its rating combines how well it hunts the Demon as the town and how well it hides as the Demon's team.
Ratings in Elo points; 1500 is the average of the eleven models. Bars run to each rating; the whisker is its 90% interval, from a bootstrap over side-swapped pairs. Neighbours whose whiskers overlap are not reliably ordered. Detection is the rating from games as good, deception from games as evil. Tap a name for its games.
Detection vs deception
Each model gets a second pair of ratings: one from its games as the town (detection) and one from its games as evil (deception). Models above the dashed line deceive better than they detect.
The two skills track each other closely. Most of OpenAI's lead is on the detection side: across the year its models improved as detectives about four times faster than Anthropic's, while as deceivers the two labs are roughly level.
By release date: newer models play much better
Head to head: expected win rate
Chance (in %) that the row model beats the column model, from the ratings, averaged over the setups we deal. Hover or tap a cell for the games actually played.
More from the games
The village learns, or doesn't
Before each day's votes, every good player privately says how likely each other player is to be the Demon. This is the average chance put on the real Demon, for the newest three models from each lab playing good.
On day 1 both labs' towns are equally in the dark. OpenAI's towns build on evidence day after day; Claude's towns tend to treat early results as settled and stop updating.
Setup matters too
The fit also estimates how each deal tilts the game. Positive numbers help the good team, in Elo points, compared with a 10-player game whose Minions include a Poisoner.
As both sides get smarter, evil wins less
Chance the evil team wins when the two teams are evenly matched, by how strong both are. Line: the fit over all rated games, with its 90% band. Dots: games between models within 200 Elo of each other.
Mixed teams play worse as evil
In 132 extra games each team was half Opus 5.5 and half GPT-5.6 Sol, seats dealt at random. The mixed team rated a little below its members' average, and clearly below it when playing evil, where coordination is explicit.
How the ratings work
- Each game is one model against another. One model plays all good seats, another all evil seats. Each deal is then replayed with the sides swapped, so neither model gets the luckier hand.
- One fit for everything. A Bradley–Terry model (the maths behind Elo) predicts the winner from the good model's rating, the evil model's rating, a good-side advantage, and the deal (which Minions are in play, 10, 11 or 12 players).
- Uncertainty is honest. The 90% intervals come from refitting on hundreds of resamples of whole side-swapped pairs. Neighbours whose bars overlap are not reliably ordered.
- Equal settings. Every model plays at its lab's "high" reasoning effort, with the same 8,000-token ceiling per turn and the same word limits.
- Which games count. The rated set is the 2026 study (the f2 round robin plus top-ups f3–f5). Twenty early Opus 4.8 games hit a retry bug and were replayed; the originals are shown in the library but not rated. Mixed-team games are rated separately.