Nvidia's AVO System Scores Perfect on ARC-AGI-3 Benchmark
Nvidia has achieved a groundbreaking milestone in artificial intelligence (AI) development. The company's AVO system, powered by Anthropic's Claude Opus 5 model, has completed the ARC-AGI-3 benchmark with a perfect score of 100%. This achievement is significant because it shows that Nvidia's agentic systems can dramatically amplify model performance.
The ARC-AGI-3 benchmark is designed to test whether AI agents can figure out unfamiliar environments without explicit instructions. The system requires agents to discover patterns, navigate unfamiliar environments, and make decisions on their own. In this case, the AVO system completed all 183 levels across 25 public environments with a reduced number of actions compared to the previous top performer, VISTA.
The underlying model powering AVO is Claude Opus 5, which manages roughly 30% on the same benchmark when running standalone. However, Nvidia's system-level architecture turned that into 100%. The company's achievement is all the more impressive because it shows that the engineering wrapped around the model, not the model itself, is responsible for the remarkable results.