Developer Andrew Boyd, founder of Standard Agents Inc., created the Jev decision model to play Pokémon Red. The model successfully completed the game in under one week. Jev is not an LLM or a chatbot; it operates by selecting options based on probability and confidence scores rather than processing visual input or generating text.
The project utilized Claude Opus 5 to monitor game logs and assist Jev when the model encountered obstacles. This collaboration involved modifying the data available to Jev and adjusting decision intervals from once per second to every six seconds to prevent loops. The changelog for the project recorded 474 entries detailing various gameplay errors and subsequent adjustments.
Human interaction also played a role, as viewers provided tips in a chat moderated by Jev. A separate implementation by Christian Mathiesen at Frigade demonstrated a different approach, with costs estimated at $1 to $1.70 per 24 hours. Another experiment by a developer named stmonty used an RTX 3080 Ti to train a world model on 42,000 frames of gameplay.
Jev launched on September 15, 2026, and saw multiple developers adopting the model within 10 days. The project highlights the potential for combining specialized decision models with LLMs for complex tasks. In contrast, an earlier attempt by Anthropic using Claude Opus 4.5 to play Pokémon Red had not reached completion as of January.
