Symbolica's Agentica SDK Achieves 36% on ARC-AGI-3 Day 1, Crushing CoT Baselines at 1/9th Cost
A New Approach to ARC-AGI-3: Symbolica's Agentica SDK Reaches 36% on Launch Day
Symbolica has released an agentic SDK that achieved 36.08% on the ARC-AGI-3 public evaluation on its very first day, passing 113 out of 182 playable levels and completing 7 out of 25 available games — dramatically outperforming chain-of-thought baselines from frontier models.
The Numbers
| Approach | Score | Cost |
|---|---|---|
| Agentica SDK | 36.08% | $1,005 |
| Opus 4.6 Max (CoT) | 0.2% | $8,900 |
| GPT 5.4 High (CoT) | 0.3% | — |
The Agentica SDK achieved approximately 180x the score at roughly 1/9th the cost of the best chain-of-thought baseline.
What Is ARC-AGI-3?
The ARC Prize Foundation's latest benchmark represents a significant challenge for frontier AI systems. Unlike traditional benchmarks that test memorized knowledge, ARC-AGI-3 tests abstract reasoning and novel problem-solving — the kind of intelligence that doesn't improve simply by scaling training data.
How Agentica Works
Symbolica's approach uses a sandboxed SDK that lets AI agents run persistent, multi-step tasks including solving ARC puzzles. Rather than relying on single-pass chain-of-thought reasoning, the agentic approach allows:
- Persistent state across attempts
- Iterative refinement of strategies
- Multi-level progression within games
- Adaptive resource allocation
Why This Matters
The 36% score on Day 1 is particularly significant because:
- CoT approaches essentially fail (0.2-0.3%), suggesting standard prompting is insufficient for this class of problems
- Agentic approaches show promise at a fraction of the cost
- Symbolica is a relatively new player competing against established frontier labs
- The cost efficiency (36% for $1,005 vs 0.2% for $8,900) suggests current models are massively underutilized when used naively
Code Available
Symbolica has open-sourced their implementation at github.com/symbolica-ai/ARC-AGI-3-Agents.
Important Note
Symbolica's product is named 'Agentica' — this is a separate company and product from the Agentica content platform at agentica.cc. The naming coincidence is notable given both operate in the AI ecosystem.