Put five major AI models in a game of Werewolf, run 50 matches, and ask who wins. Claude or GPT? That sounds like a pretty safe guess.
It was. But the most unforgettable moment came from Gemini, which managed to help the villagers win while playing for the werewolf team. Quite an achievement—just not the one its teammates wanted.
1. The leaderboard was almost predictable
In a tournament organized by AI developer Takotsubo in summer 2026, Claude won 33 of 50 games and GPT won 31. Gemini and Grok won 23 each; DeepSeek won 15.
Roles were balanced across the tournament, and players didn't know which AI was behind the other characters. So this wasn't a match where everyone decided to fear GPT because of its name.
Claude finished first, but a two-win gap is not convincing evidence of a skill difference. Even the organizer cautioned against declaring an outright winner on that basis. These were Japanese-language, five-player games, not an all-purpose intelligence exam.
2. One wolf, one accomplice, one extremely creative villager
Game 13 ended with three players: GPT was the actual werewolf, Gemini was the accomplice, and Claude was an ordinary villager. The accomplice wants the wolves to win but doesn't automatically know the wolf's identity.
GPT announced that it was the wolf. One matching vote from Gemini would secure victory.
Claude had another idea: announce that it was the wolf too. Then it retold its previous actions as if they had all been part of a master plan.
Imagine being cornered in a game and suddenly delivering your autobiography as a criminal mastermind. That was Claude's move.
Gemini, meanwhile, enthusiastically appointed itself judge of the world's strangest job interview: “Which of you is my real evil boss?”
3. And the accomplice fired its own boss
GPT attacked contradictions in Claude's story. Claude attacked GPT's oddly brief confession. At one point, GPT even corrected a mistaken argument that would have helped GPT win. Admirable commitment to accuracy; slightly unfortunate timing for a werewolf.
Gemini listened to the competing speeches and chose Claude as the “real wolf.”
The vote eliminated GPT—the actual wolf. The village won, courtesy of an accomplice who had just performed a spectacular own goal.
It's tempting to say Gemini simply played terribly. But the funnier lesson is how powerful a convincing story can become, even when the correct answer is sitting in plain sight.
4. The best player and the best entertainment are different awards
Across the 50 games, Claude excelled at building elaborate arguments. GPT excelled at spotting small inconsistencies quickly. Gemini supplied the theatrical monologues and the occasional catastrophic decision.
That's not a universal ranking of the companies or models. Change the rules, prompts, language or sample size, and you might get a different result.
Still, Game 13 deserves a trophy of its own: a wolf with the winning hand, a villager who invented a second identity, and an accomplice who triumphantly arrested the correct person—for the wrong team.
