Nvidia Reports a Perfect Public-Set Score on ARC-AGI-3 by Wrapping Claude Opus 5 in Its Own Agent

Nvidia said on August 21 that AVO, a general-purpose agent built around what the company calls Agentic Variation Operators, completed all 183 levels across the 25 public environments of ARC-AGI-3, the interactive reasoning benchmark from the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) family. AVO scored 100.00 on the benchmark's Relative Human Action Efficiency metric and, according to Nvidia, needed 12 percent fewer actions than VISTA, the previous best-performing system. The benchmark places an agent in game-like environments with no instructions, no stated rules and no declared goal, so the agent has to work out for itself what its actions do.

The detail that matters for engineers is what the harness added. Nvidia built AVO around Anthropic's Claude Opus 5, a model the ARC Prize Foundation itself measured at 30.2 percent on the same benchmark. AVO was originally designed for autonomous software engineering and graphics processing unit (GPU) kernel optimization, and the reported gap between the bare model and the wrapped model is roughly a factor of three.

Two qualifications apply to that number. Nvidia ran the evaluation on its own reimplementation of the task interface and states plainly that these results do not cover the semi-private or fully private competition sets. The ARC Prize testing policy permits this kind of self-reported public-set run, provided the reporter discloses it, and reserves official leaderboard status for runs the foundation administers itself. The public environments exist specifically so outside researchers can study them, which is also why the private sets exist.

The claim spread quickly through developer channels, including a widely shared summary of Nvidia's announcement that largely omitted the public-set caveat.