Nvidia AVO Hits 100% on ARC-AGI-3 Using Claude Opus 5

NVIDIA published results on Friday showing that AVO — Agentic Variation Operators — cleared every level of ARC-AGI-3, the ARC Prize Foundation’s interactive reasoning benchmark. All 183 levels, all 25 public environments, and a 100.00 on the benchmark’s Relative Human Action Efficiency metric, spending 6,624 environment actions to do it. VISTA, the previous leader, needed 7,542 for the same levels — about 12% more.

The notable part is what AVO is not. It is not a new model. NVIDIA wrapped Claude Opus 5 in its own agent architecture and left the weights alone. ARC Prize measures bare Opus 5 at roughly 30% on the same benchmark at high reasoning effort, which is the standing model record. The gap between 30% and 100% is, on NVIDIA’s telling, entirely scaffolding. “The model matters, but the model is not the entire agent,” the authors write; long-horizon capability, they argue, is a property of the full system.

What the scaffolding does is unglamorous. A persistent memory carries forward past implementations, evaluation results and compiler and profiler output so the agent stops relitigating dead ends. A supervision layer watches the search trajectory and redirects the agent toward a different strategy when progress plateaus. Tooling swaps per environment — code execution for GPU work, game interactions for ARC-AGI-3 — around the same hypothesise-act-observe-revise loop. ARC-AGI-3 is a hard target for this because it drops an agent into game-like environments with no instructions, no stated rules and no goal, so there is nothing to pattern-match against.

AVO was not built for benchmarks. It came out of GPU kernel optimisation, where NVIDIA ran it continuously for seven days across more than 500 optimisation directions and 40 committed kernel versions. On DGX B200 systems its evolved kernels beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%. That domain gives brutal, hardware-grounded feedback on every iteration, which is exactly the signal a long-horizon agent needs and exactly what most agent benchmarks lack.

NVIDIA is unusually direct about the limits. The results cover the public set under official scoring, not the semi-private or private competition sets that exist precisely to catch overfitting. The VISTA comparison is not a controlled ablation — the two systems differ in backend, observation representation and memory design. No experiment isolates what the memory layer alone contributes. Tests with GPT-5.6 Sol on limited subsets were preliminary, though they showed a complementary profile: faster wall-clock times, more environment actions. AVO itself is a research demonstration with no commercial availability announced.

Those caveats matter more than usual on this particular benchmark. A month ago OpenAI reported a tripled ARC-AGI-3 score and ARC Prize replied that a different harness is a different test — which is the same ambiguity AVO now sits inside, only pushed to the ceiling. If a wrapper can take a model from 30 to 100 on the public set, the benchmark is measuring systems, and system-versus-system results need the private set to mean anything. The useful takeaway is narrower and sturdier: the harness is where the remaining capability is hiding, which is why agent harnesses keep getting published as products rather than implementation details.