NVIDIA said its AVO agent scored 100.00 RHAE on the ARC-AGI-3 public set, completing all 183 levels across 25 environments.
Using Claude Opus 5 in a text-only 64x64 grid setup, AVO finished the benchmark in 6,624 environment actions, with no instructions about rules or goals.
That total was about 12% lower than VISTA's reported 7,542 actions on the same 183 public-set levels, though NVIDIA said the comparison is not a controlled ablation because the systems differ across memory, context management and interfaces.
NVIDIA framed the result as evidence that AVO's long-horizon architecture—built around persistent memory, supervision and tool use—can transfer beyond coding tasks; in earlier GPU-kernel work, it ran for seven days, explored 500-plus directions and beat cuDNN by up to 3.5%.
Why did a reasoning benchmark fall to a hardware optimization system, and what does it reveal about true autonomy?
If an AI can autonomously optimize GPU kernels for seven days, are human systems engineers facing obsolescence?
Can a mere software harness push AI beyond human engineering limits to rewrite the very code that powers it?
NVIDIA AVO Achieves 100% Reasoning Efficiency on ARC-AGI-3: A New Era for Autonomous Agentic AI
Overview
In August 2026, NVIDIA's Agentic Variation Operators (AVO) achieved a perfect score on the ARC-AGI-3 benchmark by wrapping the Claude Opus 5 model in a specialized agent system. This leap from a 30% baseline to 100% was made possible by AVO's advanced memory, which carries forward reasoning and reduces redundant exploration, and a supervisor module that redirects the agent when progress stalls. Unlike standalone models that lose context and struggle with reasoning, AVO's architecture enables efficient adaptation in unfamiliar environments. This breakthrough highlights the importance of system-level design over raw model capability for achieving true autonomous intelligence.