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

The same model the ARC Prize Foundation independently measured at 30.2 percent cleared all 183 public levels once Nvidia wrapped it in its AVO harness, a result Nvidia generated and scored on its own.

- Published: 2026-08-22T06:18:28.650Z
- Canonical: https://polylog.news/ai/2026-08-22/nvidia-reports-a-perfect-public-set-score-on-arc-agi-3-by-wr
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [NVIDIA Technical Blog](https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/), [The New Stack](https://thenewstack.io/nvidia-avo-arcagi3-benchmark/), [Polylog editors](https://polylog.news), [ARC Prize verified testing policy](https://arcprize.org/policy), [Anthropic](https://www.anthropic.com/news/claude-opus-5)

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](https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/) 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](https://thenewstack.io/nvidia-avo-arcagi3-benchmark/) 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](https://arcprize.org/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](https://t.me/aipost/7912) of Nvidia's announcement that largely omitted the public-set caveat.

## What this means

If a scaffold can triple a frontier model's score on an interactive benchmark, the thing being measured is the agent system, not the underlying model weights, and that raises the value of harness engineering relative to the value of a model subscription. Nvidia gains from that framing directly, since it sells both the computing hardware and now the agent architecture built on top of it. Model vendors lose pricing power if buyers conclude that a mid-tier model paired with a strong harness can match a premium model paired with a weak one. Anthropic is exposed on both sides of this: its model did the reasoning, and another company received the public credit for the result.

## What to watch

- Whether the ARC Prize Foundation runs AVO on its semi-private set and publishes a leaderboard figure, the only comparison that would show how much of the gain holds up outside Nvidia's own implementation.
- Whether Nvidia releases AVO to customers as a usable product rather than keeping it as a research result, which would show whether this becomes a product line or stays a demonstration.
- Whether other labs start reporting scores with the harness named alongside the model, after Meta's own coding numbers drew criticism for mixing the two together.