Nvidia AVO Scores Perfect 100 on ARC-AGI-3 Public Set

A perfect score on the hardest agentic reasoning benchmark doesn’t come from a model — it comes from what wraps around it. Nvidia reports that AVO, its Agentic Variation Operators system, completed all 183 levels across the 25 public environments of ARC-AGI-3 with a perfect 100.00 RHAE score, running on Claude Opus 5, a model that manages roughly 30% on the same benchmark by itself.

Key facts

• The result: 100.00 RHAE (Relative Human Action Efficiency) across all 25 environments in the ARC-AGI-3 public set — every one of the 183 levels completed.
• The efficiency: 6,624 environment actions versus the 7,542 that VISTA needed to clear the same levels with the same Claude Opus 5 backbone — about 12% fewer actions.
• The baseline: Claude Opus 5 alone scores about 30% on ARC-AGI-3 at high reasoning effort; the 30.2% leaderboard mark is the standing model record.
• The architecture: persistent memory that carries learnings forward, an inspect-plan-implement-evaluate loop, and a supervisor that keeps progress alive when search plateaus.

Why it matters

ARC-AGI-3 drops agents into unfamiliar game-like environments with no instructions, no stated rules, and no stated goals. Agents must infer dynamics and objectives through interaction and plan efficiently across progressively harder levels. It has become the benchmark where raw frontier models stall — which is exactly why Nvidia’s result reframes the agent debate.

“The model matters, but the model is not the entire agent,” Nvidia’s team writes. The same AVO architecture spent seven days autonomously optimizing GPU kernels before this: exploring over 500 directions, committing 40 kernel versions, and beating FlashAttention-4 performance by up to 10.5% and cuDNN by up to 3.5% on DGX B200 systems. Transferring that architecture from kernel engineering to open-ended reasoning — and saturating the benchmark — is evidence that agent scaffolding is now a first-class capability layer, not plumbing.

This lands the same month Google shipped Gemini 3.7 Flash with big agent-benchmark jumps and Z.ai’s GLM-5.3 set open-weight records on Terminal-Bench 3.0 (our coverage). The pattern across all three: evaluation is shifting from what a model scores alone to what a well-built harness extracts from it — a shift we tracked when Terminal-Bench became the agent-eval standard (explained here).

Read the fine print

Nvidia flags its own caveats, and they matter:

• Not a leaderboard score: AVO ran on Nvidia’s own reimplementation of the task interface. Official ARC Prize numbers — including Opus 5’s 30.2% — come from independently administered runs.
• Public set only: no results on ARC-AGI-3’s private held-out sets, so overfitting risk can’t be ruled out.
• Not a controlled ablation: AVO and VISTA differ in backend, observation representation, memory, and context management, so the 12% action gap can’t be attributed to memory alone.
• No product: AVO remains a research demonstration with no commercialization announced.

What to watch

Whether ARC Prize verifies any harness-assisted run on the private set — that would turn scaffolding from a caveat into the headline. And whether Nvidia productizes AVO for enterprise agent workloads, where long-horizon autonomy is the actual buying criterion.

Official source

NVIDIA Technical Blog — NVIDIA AVO Reaches 100% on ARC-AGI-3 (August 21, 2026).