Jev AI Beats Pokémon Red Using Decision Model, Not LLM

Jev, a decision model that returns solutions with a confidence figure rather than acting as a chatbot or large language model, has beaten the Elite Four and the Champion and entered the Hall of Fame in Pokémon Red. The run was completed on September 23, 2026.

Claude Opus 5 monitored the game log and adjusted the options and their wording as Jev played. The harness changelog recorded 474 entries, most pairing a failure from the game log with the change made in response.

The adjustments were not cosmetic. Jev walked into Lorelei's shut entrance 53 times and crossed one Rock Tunnel ladder 124 times in ten minutes. Opus reduced the text sent to Jev by about two-thirds and used words instead of numbers. When Jev got stuck in a loop, the harness was changed to request a decision only every six seconds instead of about once a second.

Other developers have built on the approach. Christian Mathiesen at Frigade made a second Jev-based run that reads the game's memory and lists legal options for Jev to pick from, estimating the cost at about $1–1.70 per 24 hours. A developer named stmonty trained a small world model on an RTX 3080 Ti with more than 42,000 frames of gameplay; the model picked a starter in 52 of 100 tries starting from a save in Professor Oak's lab.

Jev launched on September 15 and had multiple developers playing the game within 10 days, according to tomshardware.com.