Nvidia's harness takes Claude Opus 5 from about 30% to a perfect ARC-AGI-3 score

The number is the kind that gets repeated without its footnote, so start with both halves. On August 21 Nvidia reported that AVO, its agent architecture running on Claude Opus 5, finished every level of the ARC-AGI-3 public set. The ARC Prize baseline for that model evaluated alone, at high reasoning effort, is roughly 30%. Nvidia is also explicit that the two figures do not sit in the same experiment.

What was measured

The metric is RHAE, Relative Human Action Efficiency, which Nvidia describes as combining task completion with per-level action efficiency measured against first-time human baselines. AVO scores 100.00 on it, having completed all 183 levels across 25 environments using 6,624 environment actions.

ARC-AGI-3 is not a question-answering benchmark. Each environment is an interactive world the agent has to explore, form a theory about, and act in over many steps. That shape is why the harness matters at all: what changes between 30% and a perfect score is not the model's knowledge but how it is fed context, allowed to use tools, made to keep state, and helped to recover after a failure.

What it is compared against, and how carefully

There are two comparisons in the post and they are not equivalent. Against VISTA, another harness running on the same Claude Opus 5, AVO finished the same 183 levels using 6,624 actions to VISTA's 7,542 — about 12% fewer. That is a like-for-like comparison: same model, same levels, two harnesses.

The second comparison is the one that traveled. The roughly 30% figure is ARC Prize's evaluation of Claude Opus 5 on its own, and Nvidia says so, calling it model-level evaluation as distinct from its own system-level result. It then adds the sentence worth keeping: the comparisons should not be interpreted as a controlled ablation, because the systems differ in architecture, in how they represent what they observe, and in implementation.

So the honest reading is not that a harness turns 30% into 100%. It is that two differently built systems, one of them a bare model evaluation, produced different numbers, and nobody has isolated which part did the work.

What the paper actually covers

AVO is described in a preprint from March 2026, and its worked example is not ARC-AGI-3 at all. The paper applies the method to attention kernel optimization on Blackwell GPUs, where the discovered kernels beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on multi-head attention, and by 7.0% and 9.3% on grouped-query attention. Initial discovery took seven days; adapting it to grouped-query attention took 30 minutes.

The method replaces the fixed mutation and crossover operators of an evolutionary search with agents that consult the current lineage, a domain knowledge base and execution feedback, then propose, repair, critique and verify edits. The preprint has not been peer reviewed.

What follows from this

If you build agents, the practical lesson is that the scaffolding is a design surface with double-digit consequences, not a thin wrapper you write once and forget. Context assembly, tool access, state, and failure recovery are where the difference lives.

If you read benchmark numbers, the lesson is narrower and older: a system score and a model score are not the same measurement, and a vendor that says so in its own post has done more than most. The number to carry away is the one with the controlled comparison behind it — 12% fewer actions than another harness on the same model.