System architecture, not model alone, lifts Claude Opus 5 to perfect score on ARC-AGI-3

The gap between a language model's raw performance and its effectiveness as an autonomous agent can be enormous. That lesson emerged sharply from Nvidia's demonstration of its Agentic Variation Operators (AVO) system on the ARC-AGI-3 reasoning benchmark.

When Anthropic's Claude Opus 5 tackled ARC-AGI-3 on its own, it achieved 30.2% accuracy. Wrapped within Nvidia's AVO agent framework, the same model reached a perfect 100% score across all 25 environments and 183 levels. The finding underscores a critical insight: "System design – not model capability alone – can unlock frontier-level long-horizon performance," according to the Nvidia team.

What AVO does#

Introduced in March 2026, AVO evolved from Nvidia's work optimizing GPU kernels. The system handles the operational machinery that allows an agent to persist over long, multi-step tasks: inspecting and editing code, running commands, consulting references, and validating outcomes through execution. Its defining feature is what engineers call "sustained in-context autonomous operation" — the ability to maintain coherent progress across problems that exceed a single model's context window.

The ARC-AGI-3 benchmark measures reasoning through a metric called Relative Human Action Efficiency (RHAE), which combines task completion with the efficiency of actions taken relative to first-time human performance. It's a demanding test of adaptive problem-solving in unfamiliar environments.

Why system design matters more than model strength#

Nvidia's five-person research team, led by principal engineer Terry Chen, observed that both GPU-kernel optimization and interactive reasoning tasks follow the same underlying loop:

• Build hypotheses from incomplete information
• Take actions and observe results
• Preserve useful state and prior findings
• Refine the model of the problem
• Recover from wrong assumptions
• Continue progress across long horizons

Two mechanisms proved essential to making that loop work at scale: persistent memory and supervision.

Persistent memory allows the agent to carry forward prior attempts, evaluation results, accumulated reasoning, and learned patterns—avoiding the need to repeatedly reconstruct the search from scratch. A programmatic supervisor monitors progress and intervenes when the agent stalls, keeping work on track.