A benchmark that no frontier system could touch six months ago has just been cleared end to end — and the system that did it is not a new model. It is a wrapper around an existing one.
What Happened
On August 21, NVIDIA published results for a system it calls AVO, short for Agentic Variation Operators, showing a score of 100.00 on the public set of ARC-AGI-3, the ARC Prize Foundation's interactive reasoning benchmark. AVO cleared all 183 levels across the 25 hand-built game environments in that public set. For a benchmark where frontier systems were scoring below one percent as recently as March, the jump is not incremental.
The efficiency numbers matter as much as the headline. AVO finished the public set in 6,624 environment actions, against 7,542 for VISTA, the previous leader — roughly twelve percent fewer moves. The metric NVIDIA reports, RHAE, is scaled against what a median human manages when encountering these environments for the first time, so a score of 100.00 means the system is matching typical human efficiency rather than brute-forcing its way through with unlimited attempts.
Here is the part that reframes the whole result. The model doing the reasoning inside AVO is not NVIDIA's. It is Anthropic's Claude Opus 5, which scores somewhere around thirty percent on the same set when run on its own. NVIDIA's contribution is the architecture wrapped around it — the loop that generates candidate strategies, evaluates them against the environment, and evolves the ones that survive. NVIDIA also paired AVO with GPT-5.6 Sol on a subset of the harder games, and reported that the two models produced different operating profiles: Sol tended to reach matched levels faster in wall-clock time, while Opus used fewer environment actions.
AVO was never designed for puzzle games. It was built roughly half a year earlier for autonomous GPU kernel optimization, where it ran continuously for seven days, explored more than 500 optimization directions, and produced attention kernels that outperformed cuDNN by up to 3.5 percent and FlashAttention-4 by up to 10.5 percent on DGX B200 hardware. The accompanying paper describes the method as autonomous evolutionary search — and the ARC-AGI-3 run is essentially the same machinery pointed at a different problem class.
Why It Matters
The interesting claim buried in this result is that the ceiling on agent performance may currently sit in the scaffolding rather than the weights. A model that manages thirty percent alone and one hundred percent inside a search loop is telling you something specific: the reasoning capability was already latent, and what was missing was persistence, structured exploration, and a way to keep track of what had already been tried.
That has an economic shape to it. If capability gains can be bought with orchestration rather than pretraining runs, the barrier to entry for frontier-level agent behavior drops considerably — you no longer need to own a foundation model to compete at the top of an agentic benchmark. But it also shifts the cost from training to inference. Evolutionary search means running the model many times over, discarding most of what it produces, and keeping the survivors. NVIDIA's write-up is notably quiet on end-to-end wall-clock time and total token spend, which is exactly the number that would tell you whether this is a practical technique or an expensive demonstration.
There is also a strategic read. NVIDIA sells the compute that evolutionary search consumes. A method that turns a fixed model into a better agent by spending more inference cycles is, from NVIDIA's position, a rather convenient thing to have proven. That does not make the result less real, but it is worth naming.
Reaction
The response among practitioners has been split between genuine interest and pointed asterisks. The discussion on Hacker News surfaced the qualifications faster than most coverage did, and the first substantive comment was a correction rather than applause.
The central objection is about what ARC-AGI-3 was supposed to measure. The benchmark was designed by François Chollet's team to test whether an agent could enter an unfamiliar environment with no instructions, no stated rules, and no declared goal, and figure out what to do. Restricting harnesses was part of the point — the concern was that a benchmark-specific wrapper could present the game to the model in an artificially friendly way and hollow out the test. Several commenters argued AVO does not fall into that trap, because the evolutionary loop is general-purpose and carries no ARC-specific tooling. Others countered that any sufficiently capable harness makes the score a measurement of engineering rather than of the model's own reasoning.
The second objection is narrower and harder to argue with: this is the public set only. ARC Prize maintains semi-private and private held-out sets specifically to catch overfitting, and the verified leaderboard high sits at forty percent. One of the paper's authors joined the thread to note that the work dates back roughly six months to the kernel optimization project, and that the same evolutionary approach has since been used to generate an inference stack outperforming established serving frameworks on several tasks. Underneath all of it ran the familiar argument about whether any of this counts as general intelligence — a debate that, as one commenter dryly observed, tends to move its own goalposts.
What's Next
The obvious next step is submission to the held-out sets. Until AVO runs against the semi-private and private environments under competition conditions, the 100.00 remains a public-set number, and the gap between public and private performance is precisely the thing ARC Prize built those sets to expose.
ARC Prize 2026 is currently running with more than two million dollars in prizes and a Kaggle track built specifically around agents playing ARC-AGI-3 games, so there is a structured venue for exactly that comparison. Whether NVIDIA enters, and whether it releases or open-sources AVO — a question raised repeatedly in the community discussion and not yet answered — will shape how much the result travels. The broader open question is transferability: AVO has now been pointed at GPU kernels, inference stacks, and game environments with results in all three, which is a stronger argument for generality than any single benchmark score.
Closing Thoughts
There is a long tradition of measuring intelligence by handing a subject a grid of shapes with no instructions and seeing whether they can infer the rule. Raven's Progressive Matrices worked that way for most of the twentieth century, and ARC-AGI is a direct descendant of that idea — deliberately stripped of language, culture, and prior knowledge so that whatever remains is closer to raw induction.
What AVO complicates is the assumption that the subject taking the test is a single thing. When a model scores thirty percent alone and one hundred percent inside a loop that lets it try, fail, remember, and try differently, the question of what exactly was measured becomes genuinely hard. Humans do not solve these puzzles in one pass either; we sketch, backtrack, and abandon approaches. It may be that a search loop is less a crutch bolted onto the model than a reasonable approximation of how the reasoning was always going to have to work. Either way, the number that will settle the argument is the one from the private set, and that one has not been run yet.
한글 요약
엔비디아가 8월 21일 공개한 AVO(Agentic Variation Operators)가 ARC 프라이즈 재단의 상호작용형 추론 벤치마크 ARC-AGI-3 공개 세트에서 100.00점을 기록했습니다. 25개 게임 환경의 183개 레벨을 모두 통과했고, 소요된 환경 액션은 6,624회로 직전 1위 VISTA의 7,542회보다 약 12% 적었습니다. 불과 지난 3월만 해도 최상위 AI 시스템들의 점수가 1% 미만이었던 벤치마크입니다.
주목할 점은 AVO가 새로운 모델이 아니라는 것입니다. 실제 추론을 담당하는 모델은 앤스로픽의 클로드 오퍼스 5로, 단독으로는 같은 세트에서 약 30% 수준에 머뭅니다. 엔비디아가 만든 것은 그 모델을 감싸는 진화적 탐색 구조, 즉 후보 전략을 생성하고 평가한 뒤 살아남은 것을 발전시키는 루프입니다. 원래는 GPU 커널 최적화용으로 개발돼 DGX B200에서 7일간 연속 실행하며 cuDNN 대비 최대 3.5%, FlashAttention-4 대비 최대 10.5% 빠른 어텐션 커널을 만들어냈던 시스템입니다.
다만 이번 성적은 공개 세트에 한정되며, 과적합을 걸러내기 위한 준비공개·비공개 세트에서는 아직 검증되지 않았습니다. 공식 리더보드의 검증된 최고 점수는 40%입니다. 연구 커뮤니티에서는 모델 자체의 추론력이 아니라 하네스 설계의 성과 아니냐는 지적과, 범용 진화 루프이므로 벤치마크 전용 꼼수와는 다르다는 반론이 함께 나오고 있습니다. 총상금 200만 달러 규모의 ARC 프라이즈 2026이 진행 중인 만큼, 비공개 세트 결과가 논쟁의 실질적인 판정 기준이 될 전망입니다.
참고 — NVIDIA Technical Blog · ARC Prize — ARC-AGI-3 · arXiv:2603.24517 · Hacker News discussion