What's a person and what's a program
- Who
- Harrison Johnson. Product leader and founder, based in New York. Route 101 is a side project.
- What I decided
- The question (can a model finish a long game?), the scoring rule (the game's own story flags), the harnesses, the move and dollar budgets, what gets published and what gets cut, and every caveat on these pages.
- What I built
- A Mac build of the Pokémon Emerald decompilation with a control channel, so a program can press buttons and read the game's memory. On top of it, a Python harness: tools the models call, a memory that gets rewritten every 50 moves, seeded runs, live overlays for the videos, and the exporter that turns run logs into these pages.
- What the AI did
- A lot of the code. I work with Claude Code day to day; it wrote most of the harness and this site under my direction, and I reviewed, tested and shipped it. The questions, the experiment design, what to change between runs and what to publish are mine; Claude Code ran the runs and did the counting.
- The models
- Called through OpenRouter at each provider's default settings. The host that served a call isn't logged yet, and open models may be served by third parties in compressed form; those rows are flagged. Jev comes from TypeSafe through its public API; I have no affiliation with TypeSafe.
- The code
- Private for now. The harness lives in the same repository as the game port, and the game's source can't be redistributed. Separating the harness into a repo that can be public is on the list. Until then, every race page links its full move logs and result JSON.
- Where the idea came from
- Anthropic's Claude Plays Pokémon, which has streamed one Claude model through Pokémon Red on Twitch since February 2025. Route 101 asks a narrower question. Several models start from the same Emerald save with the same controls and move budget, the game's own story flags keep score, and every move's cost is logged.
- Scope
- Each race is one run per model on one seed, so read it as an observation. The findings change one thing at a time for one model, with six runs per setting. Still small samples, and every number links to its logs.
How a race is set up
One starting save
Every model starts from the same new game. A fixed seed locks the random numbers, trainer ID and clock. It fixes the starting point. Once two models take different paths, they roll different encounters.
One harness per race
Walk, talk, read, pick from menus. Every model in a race gets the identical toolkit and sees the same information. The findings change the controls on purpose, one thing at a time.
A fixed move budget
The same number of moves for everyone. A dollar cap only stops runaway loops; it never decides a race.
The game keeps score
Exits come from the game's own story flags. What a model claims it did doesn't count. Every move is published.
Questions
Is this made by Nintendo or The Pokémon Company?
No. Route 101 is an independent fan project about AI models. It isn't affiliated with or endorsed by Nintendo, Game Freak, Creatures or The Pokémon Company.
Does one race tell you which AI is best?
No. One race shows what happened on one save, and luck plays a part. Every result says how many runs it's based on. The findings use six runs per setting, change one thing at a time, and fix their measures before the first run.
Where can I watch the races?
Every race is on the Route 101 YouTube channel and plays on its race page. Clips also go up on X at @harrisonjohnson.
How is this different from Claude Plays Pokémon?
Claude Plays Pokémon streams one model through a long game. Route 101 runs short, repeatable tests. Several models or several setups start from the same save with the same budget, so the runs can be compared move for move.
Say hello
Questions about a run, the method or a model's row: @harrisonjohnson on X. The races go up first on YouTube; there's also an RSS feed.