On August 21, 2026, NVIDIA published a striking result on its developer blog: AVO (Agentic Variation Operators), the company’s general-purpose agent system, scored 100.00 RHAE on the public set of ARC-AGI-3, an interactive reasoning benchmark. It cleared all 25 environments and 183 levels using 6,624 environment actions. For contrast, the Claude Opus 5 model on its own manages roughly 30% on the same benchmark at high reasoning effort, per ARC Prize. The catch matters: AVO is not a newly trained NVIDIA model. It is an agent system wrapped around frontier LLMs, built for sustained autonomous work over long-horizon tasks.
What ARC-AGI-3 Tests
ARC-AGI-3 is the first interactive reasoning benchmark from the ARC Prize team, and it is built on a deliberately hostile premise. The agent gets no instructions: no rules, no goals, nothing but an environment to poke at. It has to infer how the world works through exploration and learn to beat each game within a bounded action budget. Every environment is, by design, “100% human-solvable,” and the score — RHAE — measures efficiency relative to first-time human baselines. The emphasis is on skill-acquisition efficiency and long-horizon planning under sparse feedback, not on answering static questions.
In other words, the benchmark does not test what a model knows. It tests whether the system can learn something genuinely new, from scratch, inside a tight budget of actions — much closer to what “agent” means in product roadmaps than another multiple-choice exam. That is also why a 30% score and a 100% score on it are not two points on one model curve; they measure different things entirely.
One caveat belongs up front: the perfect score covers the public set only. Semi-private and private competition sets are untouched, and public puzzles get studied to death — the fate of every benchmark eventually.
Three Design Choices Behind AVO
The blog post credits the result to system architecture rather than a model breakthrough. Three choices stand out.
First, a supervisor. Beyond the main agent loop, a separate supervisor watches the search trajectory and intervenes when progress stalls — a mechanism for escaping the loops that eat autonomous agents alive. Second, persistent memory. Implementations, results, and reasoning carry forward across context windows, so later attempts start from what earlier ones learned instead of from zero. Third, text-only observation. Each state arrives as an exact 64x64 text grid, with no images sent to the model, and actions come with no rules or goals attached.
Efficiency matters as much as the score. VISTA, another system, also reached 100% on the same 183 levels using Claude Opus 5 — but needed 7,542 actions to AVO’s 6,624, about 12% more. On a benchmark that scores efficiency against human first-play baselines, doing the same job with fewer actions is not a detail; it is the metric. There is an engineering footnote too: a seven-day attention-kernel optimization run on DGX B200 systems explored more than 500 directions and produced kernels that beat cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%.
The Model Is Not the Entire Agent
NVIDIA’s own summary is blunt: “The model matters, but the model is not the entire agent.” The same Opus 5 family that scores about 30% unaided reaches 100% inside AVO. The authors are honest that this is not a controlled ablation, but the direction is clear: long-horizon autonomy comes from the integration of memory, tools, feedback, and recovery — the whole system — not from a model’s raw reasoning score. Read that as a claim about where the next round of capability gains is hiding: less in the weights, more in the plumbing around them.
Early comparisons add a wrinkle. In preliminary subset tests, GPT-5.6 Sol was faster in wall-clock time on some games, while Opus used fewer actions. Speed and thrift are different axes in agent evaluation, and leaderboards that collapse them into one number hide real trade-offs.
What It Means for Evaluation and Builders
For teams building on AI, the lesson is about method: model evaluation and agent evaluation are different disciplines, and procurement should compare “model plus harness” as a single unit, because the same model can land a generation apart in different systems. For benchmark designers, a solved public set is a signal to rotate challenges faster and lean on private sets. And for the industry, the result puts a number on the value of the layer above the model: the distance from 30 points to 100 is engineering that lives outside the weights.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.