Symbolica Hit 36% on ARC-AGI-3 for $1,005. Frontier Models Scored Under 1% for $8,900.

The ARC Prize Foundation dropped ARC-AGI-3 on March 25 and the results broke my mental model of how AI capability scales.

Symbolica's Arcgentica framework scored 36.08% (113 of 182 playable levels, 7 of 25 games completed) using Claude Opus 4.6 as its backbone. Cost: $1,005. For comparison, raw chain-of-thought prompting on Claude Opus 4.6 scored 0.25% and cost $8,900. GPT-5.4 managed 0.26%. Gemini 3.1 Pro, the best frontier model, hit 0.37%.

Read those numbers again. The same underlying model (Opus 4.6) went from 0.25% to 36% by changing the architecture around it. Not by making the model bigger. Not by training on more data. By building a smarter harness.

Symbolica's architecture uses a top-level orchestrator that never touches the environment directly. It delegates to specialized subagents that interact with the task, then return compressed summaries back to the orchestrator. This constrains context growth (the orchestrator never drowns in details) while maintaining high-level planning (the orchestrator always sees the full picture). It's the same pattern that works in real software teams: the tech lead doesn't write every line of code, they coordinate specialists who do.

The code is open source on GitHub. Anyone can run it.

Here's why this matters beyond benchmarks. ARC-AGI-3 specifically tests skill acquisition, the ability to learn new concepts from examples and apply them to novel situations. It's not pattern matching against training data. The ARC Prize Foundation's technical report actually alleges that Gemini 3 may have memorized earlier benchmark versions, citing a reasoning chain that correctly referenced the integer-to-color mapping used in ARC tasks without being told what it was. Benchmark contamination. ARC-AGI-3 was designed to make that impossible by keeping 110 of 135 environments private and requiring interactive skill acquisition.

For builders working on multi-agent systems: the orchestrator-subagent pattern with compressed summaries is directly applicable to your work. I've been running similar architectures in my own pipelines and the context management insight is real. Agents that try to do everything in one context window fail. Agents that delegate to specialists and aggregate summaries don't. The 100x performance gap between Symbolica's approach and raw prompting is the data point that proves it.