NVIDIA Says Its AVO Agent Hit a Perfect Score on ARC-AGI-3

TL;DR

• NVIDIA claims its AVO agent architecture scored a perfect 100.00 RHAE across all 25 environments and 183 levels of the ARC-AGI-3 benchmark.
• The model doing the reasoning, Claude Opus 5, only manages 30% on its own. The rest came from the harness wrapped around it: memory, supervision, tool use, recovery.
• AVO beat a rival system called VISTA using 12% fewer environment actions, and in a separate test it out-optimized FlashAttention-4 by up to 10.5% on NVIDIA’s own DGX B200 hardware.
• The number comes from NVIDIA’s own technical blog. Nobody outside the company has published a matching result yet, and that gap matters.

A 30% Model Turned Into a Perfect Score

NVIDIA published a technical blog on August 21 laying out results for something it calls Agentic Variation Operators, or AVO. The pitch is simple and a little audacious: wrap Claude Opus 5, Anthropic’s model, inside a purpose-built agent harness, and watch it climb from a 30% baseline to a flawless 100.00 RHAE score on ARC-AGI-3, a benchmark built specifically to test whether an AI system can handle long, multi-step tasks without falling apart halfway through.

ARC-AGI-3 isn’t a trivia test. It throws 25 separate environments and 183 levels at an agent, each one demanding sustained reasoning, memory of what happened several steps back, and the ability to recover when a plan goes sideways. NVIDIA says AVO cleared every single one. On paper, that’s not a benchmark win. That’s the benchmark’s ceiling.

NVIDIA’s own framing leans hard into one idea, stated plainly in its report: system design, not model capability alone, can unlock frontier level long horizon performance. That’s the sentence the whole announcement hangs on, and it’s a deliberate shot at the assumption that bigger models automatically mean better agents.

NVIDIA also lines AVO up against a system called VISTA, which pairs Claude Opus 5 with Codex running GPT-5.6 Sol. On ARC-AGI-3, AVO didn’t just match VISTA, it needed 12% fewer actions to get through the same environments. Fewer actions means less wasted motion, fewer wrong turns, less backtracking. If that number holds up under outside testing, it’s a genuine efficiency claim, not just a completion claim.

The Harness Did the Work the Model Couldn’t

Here’s the sentence that should stick with you: a model that scores 30% alone can hit 100% wrapped in the right scaffolding. That’s not a small gap. That’s the difference between a system you’d trust with a real task and one you wouldn’t let near your calendar. And it flips a debate that’s been running in AI circles for a couple of years now, the one where bigger models and bigger context windows get treated as the only lever worth pulling.

What if the ceiling on agent performance was never really the model at all? What if it was always the scaffolding, the memory, the error recovery, the tool routing that surrounds it? Think of it like handing a brilliant but forgetful new employee a detailed notebook, a supervisor who checks their work every hour, and a clear list of who to call when something breaks. The employee’s raw talent didn’t change overnight. What changed was whether that talent could survive contact with a long, messy, multi-step job without losing the thread. That’s what AVO claims to be: the notebook, the supervisor and the phone list, built into software.

I’ve sat through enough benchmark announcements to know a perfect score should make you sit up, not just nod along. A 100% result on a benchmark designed to be hard is either a genuine breakthrough or a sign the benchmark just got solved by brute engineering, and right now there’s no way for an outsider to tell which. Big claim, thin proof.

The VISTA comparison matters because it’s the first real head-to-head NVIDIA has offered. VISTA already pairs a frontier model with a coding agent and clears the same environments. AVO’s edge isn’t that it succeeds where VISTA fails, both apparently finish. It’s that AVO does it with 12% fewer actions, which NVIDIA frames as evidence that its supervisory loop is tighter, not just luckier.

None of this comes with an independent replication yet. NVIDIA wrote the blog, ran the benchmark, and reported the number. ARC-AGI-3 is public, which means in theory anyone can check the claim, but nobody outside NVIDIA has published a matching result. That doesn’t make the number false. It just means the burden of proof hasn’t moved yet.

This Started as a Tool for Optimizing GPU Code

AVO wasn’t built for benchmark bragging rights. NVIDIA developed it first for persistent, hands-off work on software engineering and GPU kernel optimization, the kind of grinding, iterative task that usually eats up an engineer’s week. The ARC-AGI-3 run is a repurposing, proof that a system built to babysit code optimization can also babysit abstract reasoning puzzles.

And the kernel numbers are arguably the more concrete data point here, because they’re measured against named, widely used software rather than an abstract puzzle score. Left running for seven days on DGX B200 systems, AVO explored more than 500 optimization directions and committed 40 kernel versions. The result: performance up to 3.5% better than cuDNN and up to 10.5% better than FlashAttention-4, two pieces of software that plenty of AI infrastructure teams already depend on.

I’ll admit the kernel numbers grabbed me more than the puzzle score did, mostly because you can’t argue your way around a training run that’s measurably faster or slower. cuDNN and FlashAttention aren’t some side project, they’re workhorses inside a huge share of production AI training and inference stacks. Beating them, even by single-digit percentages, on hardware NVIDIA controls end to end, is genuinely interesting on its own, separate from whatever happens on ARC-AGI-3. It also happens to be the kind of result NVIDIA has every incentive to publish, since faster kernels on NVIDIA chips is a pitch the company has been making for years.

Where This Goes From Here

The obvious thing to watch is independent verification. ARC-AGI-3 is a public benchmark, so a university lab, a rival AI company, or an independent researcher could in theory run a comparable harness against the same 25 environments and see if the number holds. Until that happens, the 100% score stays a claim rather than a confirmed result.

Worth watching too: whether NVIDIA opens up AVO itself, or just keeps publishing results about it. A harness architecture is only as useful to the rest of the field as the access it gets. If AVO stays confined to NVIDIA’s own blog posts, it functions more as a marketing signal for NVIDIA hardware than as a tool anyone else can build on.

And there’s Anthropic’s side of this to watch. Claude Opus 5 is the model being wrapped here, and NVIDIA’s whole argument rests on that model’s raw 30% baseline versus its harnessed 100%. If Anthropic or an outside team runs different harnesses around the same model and gets meaningfully different numbers, that will tell you how much of AVO’s win is genuinely novel scaffolding versus tuning aimed squarely at this one benchmark.

Source: NVIDIA