- Game
- Pokémon Emerald, new game, seeded RNG, trainer ID and clock
- Seed
- 101
- Tools
- pathfinding tools (walk to a tile, talk, read, pick from menus)
- Move budget
- 300 per model
- Safety cap
- $25 per run (stops runaway loops; moves decide the race)
- Model settings
- each provider's default settings
- Score
- furthest story exit from the game's own flags, then fewest moves to reach it
- Race ends
- after 300 moves, or at the safety cap, or if the model's API fails 3 moves in a row
- Hosting
- OpenRouter default provider routing
- Cost
- cost billed by OpenRouter
- Harness
- These runs predate per-run version stamps. The code that plays the race hasn't changed since.
- Runs
episode-1.s101.claude-fable-5.1 episode-1.s101.gpt-6-astra episode-1.s101.gemini-3.1-pro-preview episode-1.s101.kimi-k3
All four models ran through OpenRouter's default routing, and the host that answered each call wasn't logged. Closed models are served by their maker or its cloud partners. Kimi K3 is an open model, so its calls may have gone to third-party hosts running compressed versions. Future races pin every model to its maker's own servers and log the host.
Moves where a model didn't act: Kimi K3 3 (2 cut off by a reply-length limit).