TypeSafe ai's Jev beats pokemon Red in under a week

Jev selects from a list of options and returns solutions with a confidence figure, but it does not read the screen or produce text or images. Anthropic’s Claude Opus 5 monitored the game log and coached Jev by adjusting options and wording. Opus reduced the text sent to Jev by about two‑thirds and changed the request frequency to every six seconds. Human viewers also sent tips that were incorporated into Jev’s option list. The harness changelog recorded 474 entries, many failures such as walking into Lorelei’s entrance 53 times.

Other runs followed. Christian Mathiesen at Frigade used Jev with a memory‑reading harness, estimating a cost of about $1–1.70 per 24 hours. A separate experiment by stmonty trained a small world model on an RTX 3080 Ti with more than 42,000 frames; it picked a starter in 52 of 100 tries. Jev launched on September 15, and LangChain described it as having an “outsized response.” Anthropic’s Claude Plays Pokémon stream, running Opus 4.5, had not finished Red as of January, but with Jev playing and Claude writing rules, the developer achieved victory.