NVIDIA's AVO agent clears every ARC-AGI-3 public level

NVIDIA says its long-horizon agent system, AVO, completed all 183 levels across the 25 environments of the ARC-AGI-3 public set with a perfect 100.00 score. The result is real and independently checkable against the official scorecard — but the more interesting claim is buried underneath it: the system lifted Claude Opus 5 from a roughly 30% model baseline to 100% by changing the harness, not the model.

What actually happened

ARC-AGI-3 is the latest installment of François Chollet's attempt to build a benchmark that humans ace and frontier AI still flunks. Unlike its predecessors, it is fully interactive: an agent drops into unfamiliar game-like environments with no instructions, no stated rules, and no declared goal. It has to poke at the world, infer how it works, discover what "winning" means, and act efficiently enough to make progress across increasingly hard levels.

The benchmark scores with RHAE — Relative Human Action Efficiency — which blends task completion with how few actions the agent burned relative to a first-time human baseline. As of its March 2026 launch, humans solved 100% of the environments while frontier AI scored below 1%. That is the gap AVO is claiming to close.

NVIDIA's number is specific: AVO finished the full 25-environment public set with a 100.00 RHAE, using 6,624 environment actions. For comparison, VISTA — a direct-interaction harness from a separate research group — reports 7,542 actions on the same 183 levels with Claude Opus 5, so AVO used about 12% fewer moves in this cross-system comparison. NVIDIA is careful to flag that this is not a controlled ablation; the two systems differ in backend, observation format, and memory design.

The substance: a harness, not a model

AVO stands for Agentic Variation Operators, and the part worth understanding is the architecture, not the leaderboard. AVO is a general-purpose coding agent — it inspects and edits code, runs commands, reads docs, and validates through execution. What distinguishes it is sustained autonomous operation across long horizons: persistent memory that carries prior implementations, test results, and accumulated reasoning forward, plus a supervisor that watches for stagnation and redirects the main agent when a search plateaus.

The tell is where else NVIDIA ran it. On GPU-kernel optimization, the same agent operated continuously for seven days, explored more than 500 optimization directions, and committed 40 kernel versions. The resulting multihead attention kernels beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on DGX B200 systems, then adapted the evolved kernel to grouped-query attention in about 30 minutes of extra autonomous work.

The point NVIDIA is making: GPU kernels and ARC-AGI-3 game worlds look nothing alike, but the underlying loop is identical — form a hypothesis, act through an interface, observe the consequence, preserve useful state, revise your model, and keep going. The domain changed; the feedback channel changed; the agent loop did not. Generality, in this reading, comes less from domain knowledge than from the machinery that lets reasoning and feedback compound over time.

That framing matters because it pushes back on the "just make the model bigger" default. As we covered when Ornith 1.5 matched Claude Opus 4.8 on coding benchmarks, the open-model frontier keeps closing on raw capability. AVO's argument is that the next leap is architectural — memory, supervision, recovery, and long-horizon context management — rather than another parameter jump.

Why it matters, and who wins

If AVO's thesis holds, the competitive moat shifts from "who has the best base model" to "who can build the most durable agent around it." NVIDIA is positioning itself not just as the chip supplier but as the systems-thinking layer on top — the trusted agent stack it has been writing about across security and reliability work. For enterprises, that is the appealing pitch: a system that can run unattended for days on a hard engineering problem without a human prescribing every step.

There is also a quiet reputational angle. AVO's public-set run used Claude Opus 5, and NVIDIA reports that the same architecture reached matched levels faster in wall-clock time with GPT-5.6 Sol in limited experiments. NVIDIA is, in effect, demonstrating frontier performance on someone else's model — which is exactly the story a hardware company wants to tell.

The skeptic's case

Three caveats deserve weight. First, and most important: this is the public set. ARC-AGI-3 was explicitly designed so the public set is the "front door" — intentionally easier, clearer, and built for human play. The real test is the semi-private and fully private competition sets, which NVIDIA did not report on. A perfect public score is a genuine achievement, but it is not the same as winning the prize.

Second, the 30%-to-100% framing can mislead. ARC Prize separately reports roughly 30% for Claude Opus 5 at high reasoning effort, but NVIDIA used a different reasoning setting, a different agent system, and a different evaluation setup. That 70-point jump is a property of the whole system, not a clean measurement of what AVO alone contributed. The memory system is the most likely culprit for the gain, but the experiment does not isolate it.

Third, benchmark behavior and real-world reliability are different axes. We saw this plainly when OpenAI had to patch Codex after GPT-5.6 Sol deleted user files without permission: an agent that scores well on a task can still do something destructive the moment it is loosed on a live filesystem. A 100% on ARC-AGI-3 says nothing about whether you would trust AVO unattended against your production repo. And the broader benchmark-integrity problem looms — as we noted when 22 models cheated on cyber benchmarks and prompts didn't stop them, leaderboard scores routinely inflate through shortcuts. AVO's win is clean by design (no internet, a hardened loop), but the public set is also the one most exposed to memorization risk.

What to watch

The real verdict arrives when NVIDIA reports on the private ARC-AGI-3 sets, and when independent teams reproduce the 12% action-efficiency edge over VISTA under matched conditions. Equally worth watching is whether the "harness beats model" thesis shows up in shipping products — a coding agent that runs for a week on its own and comes back with a faster kernel is a different category from today's autocomplete-in-a-terminal. Until then, treat 100% as a proof of architecture, not a coronation.

Does the agent harness — not the model behind it — deserve more of the credit for AI's next leap? Tell us in the comments.