NVIDIA Says AVO Took Claude Opus 5 From 30% to Perfect on ARC-AGI-3

NVIDIA says it turned Anthropic’s Claude Opus 5 from a model that completed about 30% of ARC-AGI-3 into a system with a perfect score on the benchmark’s public environments. The change did not come from a new model or altered weights, but from an agent harness that keeps the model working through a task with memory, supervision and tools.

A public-benchmark sweep

The AVO-enhanced system completed all 183 levels in ARC-AGI-3’s public environments. It used 6,624 actions, which NVIDIA’s cited comparison puts at about 12% fewer actions than the previous most efficient system. That outcome is a large gain in both completion and efficiency, rather than a score achieved simply by taking more attempts.

ARC-AGI-3 uses game-like environments with no instructions, stated rules or stated goals. Agents must infer what to do through trial and error. Its public set has 25 environments, each with six to 10 levels, and evaluates completion alongside action efficiency against people encountering the games for the first time.

What sits around the model

AVO stands for Agentic Variation Operators. NVIDIA describes it as a harness: the software layer that supplies context, maintains the agent’s state, connects tools and determines how it responds when work goes wrong. The reported benchmark improvement involved no retraining, fine-tuning or weight changes to Claude Opus 5.

• Persistent memory carries forward prior attempts, learning and reasoning so the agent can continue rather than start over.
• A supervisor monitors the main agent, intervening when progress stalls and steering it toward another strategy.
• An iterative loop repeats inspection, planning, action and evaluation, while swappable tools tailor the system to a task.

A design intended to travel

NVIDIA first built AVO to optimize its own GPU code. For ARC-AGI-3, it replaced the code-oriented tools with game controls. That sequence is central to NVIDIA’s case for the harness: the surrounding system is meant to be adapted to a new task without changing the model beneath it.

The remaining test

The result is confined to the benchmark’s public environments. ARC-AGI-3’s private competition sets remain the harder proving ground, so the perfect public result does not settle how the system will fare there. It also demonstrates performance on a structured puzzle benchmark, not a deployed enterprise workflow. The immediate significance is narrower but substantial: better orchestration around an existing model can alter the outcome of a long-running agent task without a model update.