VISTA Takes Claude to ARC-AGI-3 Perfect Score

The breakthrough is less about a smarter model than a better interaction loop: giving multimodal agents durable, revisitable visual context dramatically improves long-horizon performance.

• –VISTA archives raw visual frames so Claude can revisit evidence instead of relying on lossy summaries.
• –The result shows that agent harnesses, memory, tools, and feedback can matter as much as the underlying model.
• –ARC-AGI-3 measures action efficiency, making this a stronger systems result than a simple completion-rate benchmark.
• –The claim is limited to the public set; the authors note that model exposure to those games cannot be ruled out, leaving private-set generalization unresolved.