NVIDIA's AVO scored 100.00 RHAE on the ARC-AGI-3 public set, clearing all 183 levels across 25 environments in 6,624 environment actions

NVIDIA announced on August 21, 2026 that AVO (Agentic Variation Operators), its long-horizon autonomous agent architecture, reached a 100.00 RHAE score on the ARC-AGI-3 public set. The run cleared all 183 levels across 25 environments using 6,624 environment actions, with Claude Opus 5 as the underlying model. In the same post NVIDIA stated that these results cover the public set and are not results on the semi-private or fully private competition sets, and that the comparison should not be interpreted as a controlled ablation. ASAP breaks down what the 100.00 RHAE figure actually measures, explains why this announcement splits model leaderboards from agent leaderboards, and works out what remains once NVIDIA's own caveats are subtracted.

RHAE 100.00 measures action efficiency against human baselines, not raw accuracy

RHAE stands for Relative Human Action Efficiency, and it combines task completion with per-level action efficiency measured relative to first-time human baselines. Translating 100.00 into "100% accuracy" asserts something the source does not say. The figure carries both the fact that all 183 levels were cleared and the position of the action count used to clear them against a human's first attempt.

The design of ARC-AGI-3 is what makes such a metric necessary. Agents enter unfamiliar game-like environments without instructions, explicit rules, or a stated goal. They must infer the environment's dynamics through interaction, discover the objective itself, and then act efficiently across progressively harder levels. Because the task is to discover the goal rather than to produce a correct answer, a metric that only counts right and wrong does not capture performance.

The 6,624 action figure matters in that context. Divided across 183 levels it averages roughly 36 environment actions per level. That is an average of about 36 moves to resolve a level starting from no knowledge of the rules, and that efficiency constitutes half of the final 100.00 score.

AVO is an execution architecture with persistent memory and a supervisor, not a model

AVO is a general-purpose coding agent system developed by NVIDIA, not a newly trained model. The Agentic Variation Operators in its name refer to the fact that instead of using predetermined variation steps in evolutionary search, the agent autonomously decides at each iteration what to inspect, modify, test, and commit. The variation operators are generated by the agent's own judgment rather than fixed in a script.

The architecture is described in three layers. The first is the main agent loop, which iteratively inspects context, plans, implements changes, and evaluates results using tools. The second is persistent memory, which preserves prior implementations, evaluation results, compiler and profiler outputs, and accumulated reasoning. The third is a supervision layer, which monitors the broader search trajectory and redirects the main agent toward alternative strategies when it plateaus.

All three target the same pair of failures. One is the waste of rebuilding the search space from scratch on every iteration of a long task; the other is stagnation, where the agent loops in one direction without progress. Persistent memory lets the run resume from current state, and the supervision layer corrects course without human intervention. Running work that spans days without a person watching requires both.

NVIDIA explicitly warned against placing 30% and 100.00 side by side

The contrast most likely to be quoted from this announcement is Claude Opus 5 at approximately 30% alone versus Claude Opus 5 inside AVO at 100.00. ARC Prize separately reports approximately 30% for Claude Opus 5 at High reasoning effort, and the same model family scored 100.00 within the AVO system. It reads naturally as a claim that swapping the shell tripled the score.

NVIDIA, which drew the contrast, blocked that reading. The post states that its run used the same model family under a different reasoning setting and a substantially different agent system and evaluation setup, and that these numbers therefore should not be interpreted as a direct measurement of the performance contribution of AVO. On the VISTA comparison it says this should not be interpreted as a controlled ablation, listing agent backend, observation representation, memory, context management, and other implementation details as differing between the two systems.

The presence of those caveats defines the character of the announcement. The sentence "30% to 100.00" splices two different experiments, and at least three variables shift between them at once: reasoning-effort setting, observation format, and evaluation setup. How much the harness contributed is not settled by this release. What is settled is that the AVO configuration as a whole completed the public set.

The claim the post offers instead is narrower and sturdier: evaluating a model is not the same as evaluating an agent. That proposition holds without an ablation, and it maps exactly onto what this result demonstrates.

The gap with VISTA appears as observation format and a 12% action difference

VISTA, the comparison system, is another agent harness for ARC-AGI-3, and it cleared the same 183 levels in 7,542 environment actions. AVO's 6,624 is approximately 12% fewer.

The two systems diverge starting at the observation stage. VISTA uses rendered 512×512 PNG images and instantiates its harness with Claude Opus 5 through Claude Code or GPT-5.6 Sol through Codex. AVO uses 64×64 text grids instead of images, running NVIDIA's long-horizon agent architecture with its persistent memory, supervision, and its own execution loop.

Switching observation from images to text grids looks minor but points in a clear direction. Feeding a 64×64 grid directly as text rather than pushing 512×512 pixels through a vision encoder preserves grid coordinates and cell values without loss and raises information density per token. NVIDIA records the contrast to argue that the performance difference does not come from observation modality alone, but read the other way, it indicates that observation representation remains an open design variable for grid-world tasks.

Committing 40 kernels over a seven-day run is the evidence from outside the benchmark

The second domain NVIDIA tested AVO on is GPU kernel optimization, where the agent autonomously explored more than 500 optimization directions and committed 40 kernel versions over a seven-day continuous run. The resulting kernels outperformed cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5% on NVIDIA DGX B200, and adapting the result to grouped-query attention took roughly 30 minutes of autonomous work.

This second case is more persuasive than the first in one specific respect. ARC-AGI-3 is a benchmark, and public sets are never free of contamination and over-optimization disputes, whereas kernel optimization delivers an irrefutable feedback signal in the form of execution time at every iteration. Exploring 500 directions and committing 40 means a roughly 12-to-1 ratio of discarded to kept attempts, and sustaining that discard process for seven days without a human is a direct demonstration of the problem persistent memory and supervision were built for.

The same passage contains the number to read carefully. Up to 3.5% and up to 10.5% are maxima, not averages. Which shapes and conditions produced those maxima is not disclosed here, and kernel gains tuned to a specific shape routinely fail to transfer to others. The 30-minute grouped-query attention adaptation offsets part of this concern, but the breadth of generalization remains outside what has been published.

Model leaderboards and agent leaderboards have started to diverge

Placed against the industry trend, the position of this announcement is clear. Competition over the past several years has been aligned around raising foundation-model benchmark scores, and a single leaderboard line translated directly into product competitiveness. Here, the party that produced 100.00 is not the company that built the model but the company that built the structure wrapped around it. NVIDIA layered its own architecture on Claude Opus 5 and produced an outcome different from Anthropic's model score alone.

If this pattern holds, evaluation needs two tiers: one measuring a model's capability, and one measuring what that model can finish when wrapped in a given memory, toolset, and recovery procedure. Most public leaderboards today measure only the first, while the product a user actually encounters is determined by the second. The wider the gap between model rank and product experience grows, the more a standard for the latter is demanded.

The interpretation should not be overstated, however. The conclusion that the harness is the real protagonist would require a controlled ablation, and NVIDIA stated it did not run one. The defensible sentence right now is not that harnesses matter more than models, but that model scores alone cannot predict agent performance. The two overlap without being the same.

The reusable parts for teams outside frontier labs are memory and supervision

The two components of AVO that teams outside frontier labs can reproduce are persistent memory and the supervision layer, and neither requires training a model. What is reproducible from this announcement is not a particular model but those two structural pieces: persistent memory that preserves prior attempts, evaluation results, and tool output so the next iteration does not restart from zero, and a supervision layer that separately judges whether progress has stalled and changes direction.

Those two are exactly where most agent implementations currently break. When the context window fills, the record of prior attempts is truncated and the same failure repeats; with no dedicated stagnation detector, a human has to watch and intervene. Neither problem is solved by a better model. Both are questions of execution structure.

The case also indicates how to choose tasks. The two domains where AVO delivered were not problems with predetermined answers but problems where a machine could score an attempt immediately: level completion in the game environments, execution time in kernel optimization. Layering the same architecture onto a task with no automatic scoring signal stalls at the first question of what the supervision layer would use to judge stagnation. Any team evaluating long-horizon autonomy should secure the scoring signal before the architecture.

The open question is whether the same figure holds outside the public set

NVIDIA flagged the largest unverified item itself: these results cover the 25-environment ARC-AGI-3 public set and are not results on the semi-private or fully private competition sets. ARC Prize maintains private sets precisely to filter overfitting to public data, which makes the distinction a condition on the result's validity rather than a formality.

The second question is cost. The announcement gives 6,624 actions but not the tokens, inference calls, or dollars spent producing them. The seven-day continuous run in the kernel case suggests the resource scale this architecture presumes is not small, but a suggestion is not a figure. Cost efficiency relative to a human doing the same work cannot be judged from this release.

The third question is portability. How well AVO holds up on models other than Claude Opus 5 is undisclosed beyond limited subset experiments with GPT-5.6 Sol. If the architecture's contribution is large, performance should persist to some degree across model swaps; if it does not, this result is specific to one pairing of model and structure. That number would settle the harness debate, and it has not been published.

Three facts are established as of now. NVIDIA's AVO cleared 183 levels across 25 ARC-AGI-3 public-set environments in 6,624 actions for a 100.00 RHAE score, the underlying model was Claude Opus 5, and the result is limited to the public set and is not a controlled ablation. The rest remains open.

Source: NVIDIA Technical Blog, "NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents" (August 21, 2026), and the ARC Prize Claude Opus 5 results page cited therein. Compiled by ASAP.