Summary: A 21 August 2026 NVIDIA developer blog reported that AVO scored 100.00 RHAE on the public ARC-AGI-3 set and cleared all 183 levels in 6,624 environment actions. The stated score covers only the public set, while VISTA also reached 100% but needed 7,542 actions. NVIDIA says the result came from supervisor, persistent memory, and text-only observation rather than a model breakthrough alone.
A post on NVIDIA’s developer blog dated August 21, 2026 carried a result that is hard to ignore: AVO (Agentic Variation Operators), the company’s agent system, scored 100.00 RHAE on the public set of ARC-AGI-3, a benchmark for interactive reasoning. It cleared all 25 environments and all 183 levels in 6,624 environment actions. For scale, the Claude Opus 5 model working on its own manages roughly 30% on the same benchmark at high reasoning effort, according to ARC Prize. The qualifier here is easy to overlook and worth stating: AVO is not a model NVIDIA trained. It is an agent system wrapped around frontier LLMs and built for autonomous work that runs long.
What ARC-AGI-3 Tests
ARC-AGI-3 is the first interactive reasoning benchmark from the ARC Prize team, and its premise is deliberately unforgiving. The agent is handed nothing: no rules, no goals, only an environment to poke at. How the world works has to be inferred through exploration, and each game has to be learned well enough to finish inside a bounded action budget. By design every environment is “100% human-solvable,” and the score — RHAE — expresses efficiency relative to first-time human baselines. What counts is skill-acquisition efficiency and long-horizon planning under sparse feedback, not answers to static questions.
So the benchmark asks not what a model knows, but whether a system can pick up something genuinely new, from scratch, inside a tight budget of actions — far closer to what “agent” means in product roadmaps than to another multiple-choice exam. That is also why 30% and 100% here are not two points on one model curve; the two numbers measure different things.
One caveat belongs up front: the perfect score covers the public set only. The 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
Credit goes to system architecture rather than to a model breakthrough, and three choices stand out.
The first is a supervisor. Outside the main agent loop sits a separate supervisor that watches the search trajectory and intervenes the moment progress stalls — a mechanism for escaping the loops that eat autonomous agents alive. The second is persistent memory. Implementations, results, and reasoning accumulate across context windows, so a later attempt inherits what earlier ones found instead of starting from zero. The third is text-only observation. Every state arrives as an exact 64x64 text grid, no images sent to the model, and actions come with no rules or goals attached.
Efficiency counts as much as the score. VISTA, a separate system, also hit 100% across the same 183 levels with Claude Opus 5 — but took 7,542 actions to AVO’s 6,624, about 12% more. On a benchmark that scores efficiency against human first-play baselines, finishing the same job with fewer actions is not a detail; it is the metric. There is an engineering detail worth recording 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 pulls no punches: “The model matters, but the model is not the entire agent.” The same Opus 5 family that manages about 30% unaided reaches 100% once inside AVO. The authors say plainly that this is not a controlled ablation, yet the direction is clear: long-horizon autonomy comes from memory, tools, feedback, and recovery integrated into one 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 separate axes in agent evaluation, and leaderboards that collapse them into a single number hide real trade-offs.
What It Means for Evaluation and Builders
For teams building on AI, the lesson concerns method: model evaluation and agent evaluation are distinct disciplines, and procurement should compare “model plus harness” as one unit, since an identical model can land a generation apart depending on the system around it. For benchmark designers, a solved public set is a signal to rotate challenges faster and lean on private sets. And for the industry at large, the result attaches a number to the value of the layer above the model: the distance from 30 points to 100 is engineering that lives outside the weights.