ARC-AGI-3 on ccfarm Opus — slow tier vs fast tier
FINISHED · page rendered 2026-10-14 18:25:35 UTC
All 25 public ARC-AGI-3 games (official arc-agi toolkit 0.9.9, games run locally, scored by the toolkit's scorecard against the human baseline action counts). Model opus through https://ccfarm.chuade.com/v1/chat/completions, one access token per tier (arc-agi3-slow: native CLI agent terminals; arc-agi3-fast: direct subscription relay). The same agent plays both: each game is one growing chat (full 64×64 grid first, then the changed cells, a full grid every 10 steps), one JSON action per call, a budget of — actions per game, 4 games at a time per tier.
slow tier
not started
fast tier
not started
Per game
| Game | slow: levels | score | actions | calls / errors | mean s | $ | fast: levels | score | actions | calls / errors | mean s | $ | Fast-tier last thought |
|---|
Call ids of every model request are in each game's JSON (ccfarm Live audit: token arc-agi3-slow / arc-agi3-fast). Raw data: slow summary · fast summary.