NVIDIA’s AVO Coding Agent Hits 100 Percent on ARC-AGI-3, Up From 30 Percent for Claude Opus 5 Alone

AVO, NVIDIA’s coding agent, scored a perfect 100 percent on the public ARC-AGI-3 benchmark by wrapping Anthropic’s Claude Opus 5 in a different set of tools. On the same benchmark, Claude Opus 5 without that harness scored only 30 percent. The result suggests the scaffolding around a model, rather than the model itself, can drive large swings in benchmark performance.

AVO is described as a harness, a software wrapper, around Claude Opus 5. It originally ran in a different role, optimising CUDA GPU kernels for NVIDIA. On ARC-AGI-3 it received no rules, no prior instruction, and no stated goals. It learned by trying, observing the result of each attempt, and correcting itself.

What did AVO actually do on ARC-AGI-3?

On the ARC-AGI-3 public set, AVO cleared all 183 levels across 25 public games, a perfect score. It did so in 6,624 actions, making it 12 percent more efficient than VISTA, which needed 7,542 actions on the same set. The system also beat FlashAttention-4 by up to 10.5 percent on the kernel optimisation task it was originally built for, after running autonomously for 7 days and exploring over 500 directions.

How much of the lift came from the harness?

Claude Opus 5 on its own scored 30 percent on the same public set. AVO, built around the same underlying model, scored 100 percent. According to the figures from Wccftech, NVIDIA changed no core agent architecture. The team swapped the GPU engineering tools used for kernel work for the ARC-AGI-3 task interface and let the same setup run.

That is the key claim. The model was not replaced, retrained, or fine-tuned for the benchmark. The tool environment changed, and the score moved from 30 percent to 100 percent. The harness, the loop that lets the model act, observe, and revise, appears to be doing a large share of the work.

What does the ARC-AGI-3 caveat mean for the result?

ARC-AGI-3 does not allow external harnesses to run against its hidden private set. That means AVO’s private-set performance is unknown. The 100 percent figure applies to the public set only, and it is not yet possible to confirm how the same setup would fare on the held-out problems the benchmark keeps behind closed doors.

This is a meaningful limit. A perfect public-set score, combined with no access to the private set, leaves open the question of whether the harness generalises beyond the 25 games that were available to train against.

Why does this matter beyond ARC-AGI-3?

Public benchmark scores are often treated as a proxy for model capability. The AVO result is a reminder that the wrapper, the tools, the loop, and the way the agent is allowed to try and revise, can shift a score far more than the underlying model does. Two systems with the same model can sit at opposite ends of a leaderboard based on tooling choices alone.

For teams evaluating models, the practical reading is that benchmark numbers reflect the whole system, not just the weights inside it. Comparing models on equal footing means matching the harness, the tool set, and the action budget as well as the model name.

What is AVO built for day to day?

AVO was originally built to optimise CUDA GPU kernels. In that role it ran autonomously for 7 days, explored over 500 directions, and produced kernels that beat FlashAttention-4 by up to 10.5 percent. The ARC-AGI-3 run is a re-use of the same agent in a different domain, with the GPU engineering tools swapped out for the ARC-AGI-3 task interface. Nothing about the core agent architecture changed for that swap.

FAQ

What is NVIDIA’s AVO?

AVO is NVIDIA’s coding agent. It is a harness, a software wrapper, around Anthropic’s Claude Opus 5, originally built to optimise CUDA GPU kernels.

How much did AVO score on ARC-AGI-3?

AVO scored 100 percent on the ARC-AGI-3 public set, clearing all 183 levels across 25 public games. Claude Opus 5 on its own scored 30 percent on the same set.

Why is AVO’s ARC-AGI-3 private-set score unknown?

ARC-AGI-3 does not allow external harnesses to run against its hidden private set, so AVO’s performance on those held-out problems has not been measured.