Symbolica's Agentica SDK achieved a 36.08% score on ARC-AGI-3 on its first day, passing 113 out of 182 playable levels and completing 7 of 25 games. This significantly outperforms chain-of-thought baselines from Opus 4.6 (0.25%) and GPT-5 (0.3%), while costing $1,005 compared to $8,900 for Opus 4.6's much lower score. The code is open-sourced on GitHub.

1m read timeFrom symbolica.ai
Post cover image
Table of contents
Gallery - Games WonScore Breakdown - All GamesChat with AgenticaReferencesAppendix

Sort: