A NVIDIA-built coding agent called AVO cleared every one of the 183 levels across 25 public games on the ARC-AGI-3 benchmark, finishing at 100%. The same underlying model, Anthropic’s Claude Opus 5 running without the harness, scored 30% on the identical set, a gap produced by tooling rather than by any change to the model itself.<\/p>\n
AVO is a software harness, a wrapper that exposes tools and a feedback loop to a language model, wrapped around Claude Opus 5. NVIDIA did not change the core agent architecture. It swapped the GPU engineering tools that AVO was originally built around for the ARC-AGI-3 task interface, then let the agent run.<\/p>\n
On the ARC-AGI-3 public set the result was a perfect score: all 183 levels across 25 public games. The agent received no rules, no prior instruction, and no stated goals. It learned by trying actions, observing the results, and correcting itself.<\/p>\n
Claude Opus 5 on its own, without the AVO harness, scored 30% on the same public set. The 70-point gap between 30% and 100% comes from the wrapper: the loop that lets the model probe the environment, see what changed, and try again. Swap the harness and the same underlying model moves from a failing score to a clean sweep.<\/p>\n
AVO was originally built to optimise CUDA GPU kernels. In that earlier role it ran autonomously for 7 days, explored more than 500 design directions, and produced kernels that beat FlashAttention-4 by up to 10.5%. The same agent, redirected at a completely different problem class, picked up ARC-AGI-3 without retraining.<\/p>\n
AVO cleared the 183 levels in 6,624 actions. That is roughly 12% more efficient than VISTA, which needed 7,542 actions on the same public set. Fewer actions for a perfect score points to a tighter search, not just a longer one.<\/p>\n
ARC-AGI-3 does not allow external harnesses to run against its hidden private set, so AVO’s private-set performance is unknown. The 100% figure and the efficiency comparison both apply to the public set only.<\/p>\n
The headline lesson is that a model’s score on a benchmark is a property of the model plus its wrapper, not of the model alone. The same Claude Opus 5 that scored 30% on its own reached 100% inside AVO, with no model-side changes. For teams evaluating agent benchmarks, the harness, the tools it exposes, and the feedback loop it closes are part of the result.<\/p>\n
AVO is NVIDIA’s coding agent, a software harness wrapped around Anthropic’s Claude Opus 5. It was originally built to optimise CUDA GPU kernels and was redirected at ARC-AGI-3 by swapping its tools.<\/p>\n
AVO scored 100% on the ARC-AGI-3 public set, clearing all 183 levels across 25 public games. Claude Opus 5 without the harness scored 30% on the same set.<\/p>\n
ARC-AGI-3 does not allow external harnesses to run against its hidden private set, so AVO’s private-set performance is unknown. The 100% result applies only to the public set.<\/p>\n