NVIDIA AVO Reaches Perfect Score on ARC-AGI-3, but Only on the Public Evaluation Set

NVIDIA's AVO agent system achieved a perfect score on the public portion of ARC-AGI-3, the most widely cited interactive reasoning benchmark in frontier AI, but the result carries a caveat the company itself acknowledged: the score covers the 25-game public evaluation set, not the semi-private or private competition sets that determine prize eligibility. The same week, Jensen Huang used NVIDIA's Q2 fiscal 2027 earnings call — where the company reported $96.2 billion in revenue, up 106% year-over-year — to position AVO as evidence that agentic AI has arrived. What the numbers show is simultaneously more impressive and more limited than the headline suggests.

How AVO Turned a 30% Model Into a 100% System

AVO, which stands for Agentic Variation Operators, is a general-purpose autonomous agent architecture developed by NVIDIA's research team. The underlying model is Claude Opus 5, which ARC Prize separately reports scores approximately 30% on ARC-AGI-3 at high reasoning effort when evaluated as a standalone system. AVO's agent system left every one of the model's weights unchanged and produced a 100.00 RHAE score across all 183 levels in all 25 public-set environments.

The mechanism responsible for that jump is not a better model. It is two architectural decisions about how the model receives context and persists state over time.

The first is persistent memory. Standard model interactions are bounded by context windows: when a session ends or a window fills, prior results are lost and the agent must reconstruct its understanding from scratch. AVO's memory layer carries forward prior implementations, evaluation outcomes, compiler outputs, profiler traces, and accumulated reasoning, allowing the agent to resume from where it left off rather than restarting. In long-horizon tasks where failure often means repeating the same unproductive steps, this distinction is decisive.

The second is a supervisor component. Rather than allowing the main agent to operate indefinitely without oversight, AVO runs a parallel supervisor that monitors the full trajectory of search. When the supervisor detects stagnation — repeated cycles that are not converging on a solution — it redirects the main agent toward alternative strategies. During a separate seven-day autonomous GPU kernel optimization experiment on NVIDIA DGX B200 hardware, this supervisor-plus-memory loop sustained productive work across more than 500 optimization directions and 40 committed kernel versions, ultimately producing attention kernels that outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% in company-conducted testing on those systems, according to the AVO research paper.

In ARC-AGI-3, the agent received each environment as a 64×64 text grid — no images, no rule descriptions, no stated goals — and had to infer mechanics through interaction. The benchmark's RHAE metric combines task completion with action efficiency relative to human baselines, making brute-force approaches costly. AVO completed 183 levels in 6,624 environment actions; the VISTA system, using a similar direct-interaction approach but a different backend, required 7,542 actions for the same 183 levels with Claude Opus 5. NVIDIA notes explicitly that this is not a controlled ablation — the systems differ in memory architecture, observation representation, and context management.

The Fine Print That Changes the Score's Meaning

ARC-AGI-3 is not a single evaluation surface. It has three tiers: a public set, a semi-private set, and a private competition set. The public set is open to any agent for evaluation and is explicitly described by ARC Prize as a development and calibration resource rather than a competition score. The semi-private and private sets are withheld specifically to prevent overfitting and ensure that scores reflect genuine generalization rather than public-set memorization.

AVO's result covers only the public set. NVIDIA's post includes an editor's note — added after initial publication — clarifying precisely this distinction. ARC Prize's published methodology makes the same point: the public set is designed for iterative testing and agent development, not for determining benchmark or prize standings. Scoring 100% on the public evaluation is a meaningful demonstration of long-horizon agent capability, but it is not the measure that determines competition results.

This is the third such perfect public-set score in six weeks. The Tycho system, which uses an approach built around explicit programmatic world models, reached 100% in late July. The VISTA system reached the same score on August 5. All three used different agent architectures. All three ran on Claude Opus 5 as the underlying model — or, in VISTA's multi-model configuration, on Claude Opus 5 alongside GPT-5.6 Sol. None has published results on the semi-private or private competition sets.

The pattern is informative. The ARC-AGI-3 public set appears to have crossed a threshold where multiple well-designed agent systems can solve it, which was not true of the model-only baseline a few months ago. Whether any of these systems generalizes to the withheld competition sets — which were designed specifically to resist the kind of public-set exposure that agent systems can accumulate — remains the open empirical question.

Read more: Verification, Not Capability, Is the Bottleneck Holding Agentic AI Back

When Agent Scaffolding Outscores the Model, Who Gets the Credit?

The AVO result sharpens a question building across the agentic AI research community: if the same frontier model produces dramatically different scores depending on the surrounding architecture, what does a benchmark score actually measure?

The standard answer is that a score measures the full system, not just the model. That framing is accurate but commercially significant. A model developer releasing weights or an API does not control how those weights are wrapped. When AVO produces a 100% score using Claude Opus 5, Anthropic receives attribution for the underlying capability, but NVIDIA built the system that deployed it. The benchmark simultaneously validates both companies' value propositions — and neither the model nor the agent architecture is sufficient alone.

NVIDIA's position in this dynamic is structurally unusual: its hardware runs the agent systems, its research team built one of them, and its Q2 FY27 earnings — $96.2 billion in revenue, $89.0 billion from data center customers — reflect the infrastructure bet that agent workloads will require far more compute than prior AI generations. Jensen Huang's earnings-call framing of AGI was deliberately terse and definition-free. His statement — "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable. Now, compute is revenue" — is consistent with NVIDIA's long-standing posture: sell infrastructure, avoid picking which lab's model wins, and claim each benchmark as confirmation that agentic AI is real, useful, and compute-intensive.

What Comes Next: Astra, Anthropic EAPs, and the Recursive Loop

The broader context for NVIDIA's AVO result is a frontier model landscape in visible motion. OpenAI's Astra, the model family internally positioned as GPT-6 and described by The Information as designed for long-running parallel agentic tasks, remained in late-stage internal testing as of this writing, following a delay introduced when the system triggered OpenAI's Preparedness Framework's Critical-tier cybersecurity evaluation in early August. OpenAI has not confirmed a release date. In a podcast interview published in late August, CEO Sam Altman described the preceding months as unusually difficult for the company and indicated that broad Astra availability remained the top near-term priority.

At Anthropic, two food-named early-access model strings — claude-marshmallow-eap and claude-melon-eap — appeared in developer tooling on August 24, following a pattern the company established before previous launches. Community evaluations suggested marshmallow outperforms melon on most tasks and that both sit below Fable 5 in overall capability, though neither result represents a controlled independent benchmark. Anthropic issued no official statement on either codename.

Read more: Anthropic Drops Two Mystery Claude EAPs as OpenAI's Astra Surfaces Another Checkpoint

Both companies are navigating these releases against IPO timelines under scrutiny. Reuters reported that OpenAI was laying the groundwork for a public offering that could value the company at up to $1 trillion, with CFO Sarah Friar having told some associates the company was targeting a 2027 listing and some advisers projecting it could come as early as late 2026; OpenAI said at the time that an IPO was not its focus and that it had not set a date. Anthropic's own IPO preparations have proceeded on a parallel track.

The deeper mechanism connecting AVO, Astra, and the model race is one both Jensen Huang and Sam Altman have articulated separately: agents capable of accelerating AI research itself. NVIDIA's AVO architecture was first validated against GPU kernel optimization — a task requiring autonomous inspection, hypothesis generation, hardware-grounded testing, and revision over seven days without human intervention. If the same loop applies to model training and architecture search, each generation could be developed faster than the last. That remains speculative as a near-term claim. But the AVO kernel result — hardware-grounded, with measurable improvements against cuDNN and FlashAttention-4 — is the closest existing demonstration that the loop is not merely theoretical.

The ARC-AGI-3 public set has now been solved three times in six weeks by three different agent systems, all using frontier models as their core. The agents that did it share a common structure: they persist what they learn, they recognize when they are stuck, and they change course. Whether that structure will transfer to the withheld private competition sets — designed specifically to resist that kind of accumulated exposure — is the question ARC Prize will eventually answer. For now, the score is real, the fine print matters, and the architecture doing the work is the agent system, not the model alone.