An artificial intelligence decision-making model named Jev has beaten Nintendo’s classic Game Boy title Pokémon Red in 37 hours and 40 minutes, according to developer Christian Mathiesen. The run cost a total of $1.65 in API usage fees while executing 16,150 distinct decisions across 1,660 battles, though the software required a specialized auxiliary framework to escape the opening town.
How the Decision Model Processed Game Data
Jev was developed by TypeSafe AI as a specialized decision model rather than a traditional text-generating large language model. The system processes inputs at a cost of $0.042 per million tokens with free outputs, selecting actions from pre-determined sets rather than generating open-ended strings. Frigade CTO Christian Mathiesen set up the experiment to test how the model handled high-frequency choices in a structured environment.
Running the game required more than raw model weights. A dedicated companion program read RAM straight from the Game Boy emulator to track party status, map locations, and combat conditions. This wrapper script fed Jev concrete options like healing, advancing, or catching monsters, while pathfinding relied on the A* routing algorithm to walk through maps once the AI picked a destination. Without these supporting rails, Mathiesen noted that the model could not move past Pallet Town on its own.
Overcoming Combat Hurdles and Technical Bottlenecks
The run saw the AI party black out 16 times and face the Elite Four 15 sessions in a row, with Victory Road proving to be the hardest block. Jev used roughly 39.2 million input tokens across the playthrough. The final hall of fame team featured an LVL 83 Charizard, LVL 62 Graveler, LVL 45 Nidoqueen, LVL 44 Beedrill, LVL 39 Haunter, and an LVL 29 Primeape.
To prevent infinite loops where the software repeated failed actions, the harness tracked unmoving game states to force alternative paths. If recovery failed, the program rolled back to the latest checkpoint. Mathiesen published the project repository on GitHub for public testing, though users must supply their own game ROM.
