Symbolica Hits 36% on ARC-AGI-3 Day 1 Using Orchestrator-Subagent Architecture for $1,005

Symbolica AI·high signal

Symbolica published their Arcgentica framework on March 25, achieving 36.08% on the brand-new ARC-AGI-3 benchmark (113/182 playable levels, 7/25 games completed) using Claude Opus 4.6 at a cost of $1,005. For comparison, baseline chain-of-thought prompting scored 0.25% (Opus 4.6 Max, $8,900) and 0.3% (GPT 5.4 High). The architecture uses a top-level orchestrator that never interacts with the environment directly, instead delegating to specialized subagents that return compressed summaries — constraining context growth while maintaining high-level planning. Code is open source on GitHub.