Opus 5.5 picked Squirtle because of Brock. Fable 5.1 picked Charmander because it was the closest ball. One cost $8, the other $100.

A benchmark tested AI models in Pokémon Red, scoring them on milestones within 1,000 turns to defeat Brock. Opus 5.5 chose Squirtle, while Fable 5.1 picked Charmander, which was more expensive. Models received a prompt, screenshots, game state, and on-screen text, but no pathfinding or map hints. One model, after repeatedly losing, decided to catch a Mankey but instead ground in Viridian Forest, evolved Charmeleon on turn 511, and beat Brock on turn 554. The overall goal was simply to beat the game, without specific instructions on turn limits or Brock as the endpoint.

Unlike other benchmarks that focus on raw performance, this test uniquely evaluates AI models on their ability to plan ahead without explicit instructions.