NVIDIA's general-purpose coding agent system AVO scored 100% across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels1,2,3. The result elevates Anthropic's Opus 5 from a 30.2% model baseline to a perfect score when wrapped in the AVO agent architecture.
A five-person NVIDIA team — Terry Chen, Jean-Francois Puget, Humphrey Shi, Yeyin (Eva) Zhu, and Zhifan Ye — described the architecture in a blog post published on Friday. The team wrote that the result "shows that system design not model capability alone can unlock frontier-level long-horizon performance".
NVIDIA first introduced Agentic Variation Operators (AVO) in late March 2026, initially demonstrating it on software engineering and GPU-kernel optimization tasks. AVO handles agentic architecture tasks including inspecting and editing code, running commands, consulting documentation, and validating its work through execution. Its distinguishing focus is sustained in-context autonomous operation across extended, multistep tasks on long horizons. AVO replaces the predefined variation step of conventional evolutionary-search systems with an autonomous agent that decides what to inspect, change, test, and commit.
The system requires persistent memory and supervision to preserve state and progress on long-horizon tasks beyond a single model's context window. AVO achieved a 100.00 RHAE score across all 25 environments using the official scorecard and RHAE metric on the public set — not the semi-private or fully private competition sets.
The 30.2% baseline for Opus 5 at high reasoning effort on the ARC-AGI-3 public set was reported by non-profit AI research and benchmarking body ARC Prize in an analysis post in July. AVO operated with no instructions, explicit rules, or stated goals across the 25 environments.
The NVIDIA team also experimented with pairing AVO with GPT-5.6 Sol on a challenging subset of games. According to the team, Sol reached matched levels faster in wall-clock time in several cases, while Opus 5 used fewer environment actions in matched-level comparisons.
ANALYSIS The 30.2%-to-100% jump on a fixed model underscores the degree to which agent scaffolding — rather than raw model capability — can determine benchmark outcomes. The head-to-head notes on Sol versus Opus 5 within the same harness offer a rare controlled comparison of two frontier models on identical agentic tasks, with each showing a different efficiency profile.