TypeSafe AI's Jev Decision Model Completes Pokémon Red in a Week With Claude's Help

TypeSafe AI's Jev reached the Hall of Fame in Pokémon Red on September 23, 2026, after beating the Elite Four and the Champion, according to the project page. The accomplishment stands out because Jev operates differently from chatbots that have taken weeks or months to complete the Blue version: it selects from a predefined set of choices rather than generating responses freely. The run succeeded with significant intervention from Anthropic's Claude Opus 5, which monitored gameplay logs and refined the available options to guide Jev forward.

Andrew Boyd, founder of Standard Agents Inc., initiated the project and announced it on X with the goal of finishing within a week. The gameplay unfolded as a livestream accessible through a browser or terminal, complete with a chat interface that Jev moderated.

Let's go! Jev Plays Pokemon. Follow along here: https://t.co/64naxTJlDg OR, in your terminal run `npx jev-plays-pokemon` to follow along (with chat!) in a TUI. Github oAuth required to chat. Jev is the player and the chat moderator. Let's catch them all!

Andrew Boyd, September 17, 2026

How Jev Works

Jev is a decision model distinct from both chatbots and large language models. Rather than generating text or analyzing screen images, it evaluates a curated list of options paired with supporting facts and selects the most probable choice based on confidence scoring. Claude Opus 5 functioned as an indirect coach, modifying the options and data presented to Jev when the model encountered obstacles, rather than directly controlling gameplay.

Challenges and Adjustments

The project's changelog documented 474 entries, predominantly recording failures from the game log alongside corresponding fixes. Notable stumbling blocks included walking into Lorelei's closed entrance 53 times, crossing a single Rock Tunnel ladder 124 times within ten minutes, and losing to the Champion's Alakazam after defeating all four Elite Four trainers, which necessitated repeating those battles before finally overcoming the Champion later that day.

Claude Opus 5 made refinements targeting both performance and expense. Drawing on TypeSafe's guidance, it cut the text volume sent to Jev by roughly two-thirds and substituted words for numerical values. When Jev entered repetitive loops, the system was reconfigured to request decisions every six seconds instead of approximately once per second. The community contributed as well: helpful suggestions from livestream viewers were incorporated into Jev's option list. This combination of machine assistance, developer intervention, and audience input highlighted both the potential and constraints of the decision model approach.

Alternative Approaches

Christian Mathiesen at Frigade conducted a separate Jev-based experiment using a different architecture. His harness, per its README, "reads the game's memory, lists the legal options ... and Jev picks one," without writing to game memory. Mathiesen's initial iteration, where Jev could directly select buttons, "never left Pallet Town." Operating costs for his version came to approximately $1–1.70 per 24 hours.

A third Pokémon Red project pursued an entirely separate direction. A developer using the handle stmonty trained a compact world model on an RTX 3080 Ti using over 42,000 frames of recorded gameplay. Beginning from a save inside Professor Oak's lab, it successfully selected a starter Pokémon in 52 of 100 attempts, according to stmonty's blog. This effort required substantially less support than Jev's run: the model had to deduce what each input accomplished solely from visual information.

Broader Impact

Jev's viral reception prompted LangChain to note it had "had a pretty outsized response" following its September 15 launch, spurring multiple developers to attempt the game within ten days. The victory demonstrates that pairing specialized models with collaborative refinement can substantially enhance problem-solving capabilities. TypeSafe has noted that open-ended tasks remain better suited to LLMs, as previously reported at Jev's introduction. By contrast, Anthropic's Claude Plays Pokémon stream, which ran Opus 4.5 at the time, had not completed Red as of January. With Jev handling gameplay decisions and Claude managing the rule adjustments, the developer achieved what the earlier approach had not.