The 70-Point Gap That Changes the Conversation

When ARC-AGI-3 launched in March 2026, the results were humbling: Claude Opus 5 scored 25%, GPT-5.4 hit 26%, and the full leaderboard sat well below 1% on a benchmark where untrained humans score 100%. Fast-forward to August 21, 2026, and NVIDIA announced that Claude Opus 5 — the exact same model — achieved a perfect 100.00 RHAE score across all 183 levels of the ARC-AGI-3 public set. The difference was not a new model. It was the harness around it.

NVIDIA’s Agentic Variation Operators (AVO) architecture, published by Terry Chen, Eva Zhu, Zhifan Ye, Jean-Francois Puget, and Humphrey Shi at NVIDIA Research, lifts a frontier model from a 30% standalone baseline to perfect public-set performance — not by changing a single weight, but by redesigning the scaffolding that surrounds it. That gap is the most important data point in the paper.

What AVO Actually Is

AVO is a general-purpose coding agent system, not a fine-tuned model or a domain-specific solver. At its core, it is an iterative loop: the main agent inspects context, plans a change, implements it, evaluates the result, updates its internal state, and continues. This is familiar territory for anyone who has worked with modern coding agents. What distinguishes AVO is what happens at the boundaries of that loop.

Two mechanisms carry the load. First, persistent memory: prior implementations, evaluation results, compiler outputs, and accumulated reasoning are retained across iterations, so the agent resumes from the current state rather than reconstructing it from scratch on each pass. Second, a supervisor monitors the broader search trajectory for stagnation — when the main agent gets stuck in repeated unproductive cycles, the supervisor redirects it toward alternative strategies. Memory prevents regressive loops. The supervisor prevents the kind of tunnel vision that kills long-horizon tasks.

Crucially, the model receives no special treatment. AVO connected Claude Opus 5 to the ARC-AGI-3 task interface in text-only mode: each observation was supplied as a 64×64 text grid, with no images or image tokens. The agent was given the list of available actions but no rules, no objectives, no instructions. It had to infer all of that through interaction — exactly as human first-timers would.

Where AVO Came From: Seven Days on a GPU

AVO was not designed for ARC-AGI-3. It was built for GPU kernel optimization — one of the most unforgiving engineering domains, where small changes in memory layout or scheduling can silently break correctness while marginally improving throughput. NVIDIA’s attention-kernel study ran AVO continuously for seven days on DGX B200 systems, exploring over 500 optimization directions and producing 40 committed kernel versions.

The results were commercially meaningful: the evolved multihead attention kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% across tested configurations. For a company whose customers pay per GPU hour, those percentages matter. But the more interesting outcome was architectural: the same agent loop — inspect, hypothesize, act, observe, revise — that worked on compilers and profilers also transferred to a completely different environment with no modification to the agent backend.

That transfer is the claim worth examining. ARC-AGI-3 has nothing in common with GPU kernels on the surface. But the underlying computational pattern — build a hypothesis from incomplete evidence, take an action, observe consequences, preserve useful state, recover from wrong assumptions, continue over a long horizon — is identical. NVIDIA’s assertion is that long-horizon capability is a property of the machinery for sustained progress, not of domain knowledge.

What the Numbers Actually Show — and What They Don’t

AVO completed the 25-environment public set using 6,624 environment actions. The previous leader, VISTA (which also used Claude Opus 5 as its base model), required 7,542 actions to complete the same 183 levels — roughly 12% more. That efficiency gain is real, but the paper is careful about what it proves: AVO and VISTA differ in agent backend, observation representation, memory design, and context management. This is a cross-system comparison, not a controlled ablation of any single variable.

More importantly, the 100% result covers only the 25-environment ARC-AGI-3 public set. The ARC Prize semi-private and fully private competition sets — the ones that determine prize eligibility — remain untested. The $2M Grand Prize requires 100% on the full evaluation. NVIDIA has not submitted for prize consideration, and the NVIDIA team is explicit about this distinction. Treating the public-set result as a solved-AGI headline would be exactly the kind of benchmark overinterpretation that makes the AI community’s signal-to-noise ratio so poor.

The RHAE metric itself is worth understanding: Relative Human Action Efficiency measures how many actions an agent takes to complete each level relative to second-best human performance on the same level, then aggregates across all environments. A 100.00 RHAE means AVO completed every level with at least the efficiency of the second-best human first-timer. For the public set, that is a genuine technical result. It is also a result on 25 known environments — not on the held-out environments that define the competition.

What This Means for Agent Design

The more durable lesson from AVO is not about ARC-AGI-3. It is about how to think about building agents for long-horizon tasks. If the same harness can transfer from GPU optimization to interactive reasoning with no model change, then the following design principles are likely to matter in production agentic systems:

Memory architecture is not optional. The AVO memory system carries forward useful understanding to reduce repeated exploration. Without it, each model invocation starts from scratch, and long-horizon performance degrades quickly. This is the single biggest gap in most deployed agent systems today — context windows are not a substitute for structured persistent memory.

Supervisory loops prevent stagnation. A main agent will optimize locally. A supervisor that can observe the broader trajectory and intervene when progress stalls is the difference between a seven-day optimization run and a seven-minute loop that fails on the same approach repeatedly. The supervisor in AVO does not do the work — it watches the work and corrects the direction.

Evaluating the model is not the same as evaluating the agent. This is the AVO paper’s central provocation. Model-level benchmarks — isolated prompts, single-turn reasoning tasks — are useful for measuring capability, but they do not predict how a model will perform inside a complete agentic system. A 30% standalone score and a 100% system score from the same model, in the same evaluation environment, should give every enterprise team pause before they select a model based on leaderboard performance alone. For context, vortx.ch covered this dynamic from a different angle in ARC-AGI-3 Milestone 1: What the Prize Winners Reveal — the pattern is consistent.

The practical implication for teams building on frontier models is uncomfortable: you may be underselling your current model and overspending on the next one. Investing in the agent architecture around a strong-enough base model may yield more performance per dollar than upgrading from Sonnet to Opus, or from one vendor to another. The Claude Opus 5 analysis from July made a similar point about cost-efficiency tradeoffs at the model layer; AVO suggests the harness layer is where the real leverage sits.

What’s Next

NVIDIA’s preliminary experiments paired AVO with GPT-5.6 Sol on a challenging subset of ARC-AGI-3 environments. The early data suggests complementary profiles: Sol completed matched levels faster in wall-clock time in several cases, while Opus used fewer environment actions. NVIDIA describes this as “preliminary” and explicitly defers a systematic comparison to future work — which is the right call. Cross-model comparisons inside a complex harness are difficult to interpret without careful controls.

The harder question is whether AVO’s public-set 100% will survive the private set. History on ARC benchmarks is not encouraging: systems that dominate the public set often find that the private environments expose generalization failures. The ARC Prize Foundation designed the evaluation structure precisely because of how quickly agents learn to overfit to known puzzle types. If AVO transfers as cleanly to the private environments as it did from GPU kernels to interactive reasoning, that would be a significant result. The community will be watching.

Further Reading

• NVIDIA Technical Blog: AVO Reaches 100% on ARC-AGI-3 — the primary source, with architectural diagrams and detailed methodology notes from the NVIDIA Research team.
• AVO: Agentic Variation Operators for Autonomous Evolutionary Search (arXiv) — the original March 2026 paper introducing AVO in the context of GPU-kernel optimization; the ARC-AGI-3 work builds directly on this architecture.
• ARC-AGI-3 Benchmark and Scoring Methodology — the ARC Prize Foundation’s documentation on the RHAE metric, environment design, and competition rules; essential context for interpreting any ARC-AGI-3 result.