NVIDIA says an AI agent system it built, called AVO, has completed every level in the public portion of ARC-AGI-3, one of the toughest benchmarks designed to test genuine reasoning rather than memorized knowledge. According to a technical blog post NVIDIA published on August 21, 2026, AVO paired with Anthropic’s Claude Opus 5 achieved a perfect score across all 183 levels in the benchmark’s 25 public environments.
That’s a striking jump. On its own, without any surrounding agent system, Claude Opus 5 scored around 30% on ARC-AGI-3 — itself a record at the time. NVIDIA’s claim is that wrapping the same model in its AVO architecture pushed that number to 100%, at least on the portion of the benchmark that’s publicly accessible.
The catch, and it’s an important one: ARC-AGI-3 keeps most of its test environments hidden specifically so that AI systems can’t be tuned or optimized against them. NVIDIA’s 100% figure covers only the smaller, open subset. The larger, private sets used to verify genuine leaderboard standing haven’t been touched by this result — at least not according to anything NVIDIA or independent observers have published so far.
Table of Contents
What NVIDIA Announced
NVIDIA’s developer blog frames the result plainly: a general-purpose agent system, not a new model, solved every level of ARC-AGI-3’s public test set. The company reports a perfect RHAE score — the benchmark’s efficiency-adjusted measure of how well an agent solves tasks relative to a human baseline — across all 183 levels spread over 25 environments.
This is, for now, a company-reported result. It appears on NVIDIA’s own blog and its official social accounts, but it has not yet shown up as a verified entry on ARC Prize’s official leaderboard, and no independent confirmation of the score is currently available.
What AVO Actually Is
AVO stands for Agentic Variation Operators, and despite the acronym showing up in headlines about ARC-AGI-3, it wasn’t built for that benchmark at all. NVIDIA first introduced AVO in a research paper posted to arXiv in March 2026, in a completely different context: optimizing GPU code.
In that original work, AVO ran autonomously for seven days on NVIDIA’s Blackwell B200 GPUs, searching for better versions of attention kernels — the low-level code that powers how AI models process information efficiently. The system reportedly found kernels that outperformed NVIDIA’s own cuDNN library by up to 3.5% and beat FlashAttention-4, a widely used optimization technique, by up to 10.5%, with those gains carrying over to a related architecture after only about 30 minutes of additional adjustment.
The ARC-AGI-3 experiment reuses that same system on a very different kind of problem — reuse that’s central to NVIDIA’s argument that AVO is a general-purpose architecture, not a benchmark-specific trick.
It’s worth being precise about what AVO is not: a new NVIDIA language model. The 100% result came from combining AVO’s scaffolding — memory, planning, tool use — with Anthropic’s Claude Opus 5 doing the actual reasoning. NVIDIA also tested AVO with OpenAI’s GPT-5.6 “Sol” on a subset of tasks, though that comparison was limited rather than systematic.
Timeline
- March 2026: NVIDIA introduces AVO in a research paper, applying it to GPU-kernel optimization.
- March 25, 2026: ARC-AGI-3 launches publicly.
- Mid-2026: Claude Opus 5 sets a raw-model record of roughly 30% on ARC-AGI-3, with no surrounding agent system.
- August 21, 2026: NVIDIA publishes AVO’s 100% result on ARC-AGI-3’s public set.
Why ARC-AGI-3 Is a Hard Benchmark
ARC-AGI-3 continues a benchmark series designed by researcher François Chollet to measure something narrower and harder than most AI tests: the ability to solve genuinely novel problems, rather than problems that resemble something in a model’s training data.
Unlike a multiple-choice quiz, ARC-AGI-3 presents interactive puzzle-like environments that an agent has to explore and figure out from scratch, without being told the rules. Humans tend to solve these easily. AI models, historically, have not. When the benchmark launched, frontier models scored roughly 0.5% or lower, compared to close to 100% for humans.
That gap is what makes any high score newsworthy — and also what makes the details of how a score was achieved so important. A model solving these puzzles unaided is a different achievement than a model solving them with the help of an elaborate support system that manages its memory, retries its failures, and keeps it on track.
The Public-Set vs. Private-Set Caveat
This is the detail that got the most attention within hours of NVIDIA’s announcement. ARC-AGI-3 splits its test environments into three groups: a public set anyone can access, a semi-private set used for controlled evaluation, and a fully private set reserved for competition scoring. In total, 110 of the benchmark’s 135 environments are held back from public view, precisely so systems can’t be shaped around known answers.
NVIDIA’s 100% score applies only to the 25 environments in the public set. Commentary on Hacker News following the announcement raised a specific concern: because AVO’s testing happened where the puzzles were visible, there’s no guarantee the same performance would hold on the private sets. Commenters also noted that the underlying achievement belongs largely to Claude Opus 5’s reasoning, with AVO functioning as a harness around it rather than a new source of intelligence in its own right.
ARC-AGI-3’s Public vs. Private Sets
- Public set: 25 environments, openly accessible — this is what NVIDIA’s 100% claim covers.
- Semi-private and private sets: 110 environments combined, held back specifically to prevent systems from being tuned to known answers and to verify genuine competitive standing.
- NVIDIA’s reported score has not been tested against the 110 held-back environments.
None of this means the result is illegitimate. It means the 100% figure should be read as “perfect on the visible portion of the test,” not as a fully verified benchmark milestone.
How AVO’s Architecture Works
NVIDIA describes AVO as a repeating loop: the system inspects its current context, plans a next step, acts using available tools, and evaluates the outcome — then repeats, carrying memory forward between steps. A separate supervisor component watches the broader trajectory of the task and can step in if the agent stalls or starts repeating unproductive patterns.
For the ARC-AGI-3 tests, the model didn’t see images of the puzzles. Instead, each environment was represented as a 64-by-64 grid of text, with no upfront instructions about the rules — matching the benchmark’s design, which tests adaptation rather than instruction-following.
NVIDIA says its interface was informed by an earlier system called VISTA, though the two aren’t identical — VISTA runs Claude Opus 5 through Anthropic’s Claude Code tool, or GPT-5.6 through OpenAI’s Codex, and in its main configuration relies on rendered images rather than text. NVIDIA built its own version of the task interface rather than adopting VISTA’s directly.
Comparison to Other ARC-AGI-3 Systems
Beyond VISTA, NVIDIA references a third system, Tycho, which takes a different approach: rather than reacting directly to each puzzle, Tycho builds an explicit, executable model of how each environment works. NVIDIA opted not to follow that path, testing AVO instead as a general-purpose agent without an ARC-specific reasoning layer.
| System | Backend Model(s) | How It “Sees” the Puzzle | Approach |
|---|---|---|---|
| AVO (NVIDIA) | Claude Opus 5 (primary); GPT-5.6 “Sol” (limited tests) | Text-based 64×64 grid | General-purpose harness: iterative plan–act–evaluate loop with persistent memory and a supervisor |
| VISTA | Claude Opus 5 via Claude Code; GPT-5.6 via Codex | Primarily rendered image | Direct interaction; informed AVO’s interface design but wasn’t adopted directly |
| Tycho | Not publicly disclosed | Not publicly disclosed | Builds an explicit, executable model of how each environment works |
On efficiency, NVIDIA reports its system used about 12% fewer actions than VISTA on the same 183 public levels — 6,624 versus 7,542. NVIDIA caveats this itself, noting the two systems differ in model access, observation format, and memory management, so the difference isn’t from a controlled, apples-to-apples test. NVIDIA also hasn’t published wall-clock time or compute cost for either system.
Is AVO Available to Developers?
No. AVO is, as far as anything published shows, an internal NVIDIA research system — not a product, SDK, or open-source release, with no code repository or developer documentation. Outside commentary describes it, alongside comparable systems like Google DeepMind’s AlphaEvolve, as closed: impressive in its reported results, but not something the public can inspect or run. NVIDIA hasn’t said whether that will change.
Why This Matters for AI Agents Generally
NVIDIA’s broader argument is that the software wrapped around a language model matters as much as the model itself — that memory management, error recovery, and task supervision can unlock capability the underlying model doesn’t show on its own. The company points to AVO’s reuse across GPU-kernel optimization and abstract puzzle-solving as evidence the architecture is genuinely general-purpose rather than tuned to one task.
That’s an active question across the industry, not one NVIDIA is exploring alone — similar work at OpenAI, Google DeepMind, Microsoft, and Anthropic addresses how to structure agents that operate reliably over long, multi-step tasks. AVO’s results add a data point to that conversation; they don’t settle it.
Limitations and Open Questions
Several things about this result remain unverified or undisclosed:
- Whether AVO’s performance holds on ARC-AGI-3’s private and semi-private environments, which haven’t been tested publicly.
- The time and compute cost of the evaluation, which NVIDIA hasn’t disclosed.
- Whether ARC Prize, the organization behind the benchmark, or François Chollet himself has reviewed or commented on this specific result — no such statement has surfaced as of this writing.
- Which organizations built the comparison systems VISTA and Tycho; that information isn’t available in current public sources.
- Any direct comparison between AVO and other well-known agent frameworks, such as Microsoft’s AutoGen or Anthropic’s Computer Use tools — no such comparison has been published.
Conclusion
NVIDIA’s AVO system, combined with Claude Opus 5, cleared every level of ARC-AGI-3’s public test set — a meaningful jump from the roughly 30% the same model manages without that support. It’s a genuine demonstration that the systems built around a language model can dramatically change what it’s able to do on a hard reasoning benchmark.
What it isn’t, at least not yet, is an independently verified benchmark milestone. The claim rests on the portion of ARC-AGI-3 that’s publicly visible, not the private sets that actually determine competitive standing, and no outside body has confirmed the score. Whether AVO’s performance survives contact with the hidden test environments — and whether NVIDIA or the ARC Prize Foundation says anything further about it — is the detail worth watching next.
Frequently Asked Questions
Did NVIDIA’s AVO really score 100% on ARC-AGI-3? NVIDIA reports a perfect score, but only on the benchmark’s public set of 25 environments. The private and semi-private sets used for official leaderboard verification haven’t been tested with this result, and no independent group has confirmed it.
What is NVIDIA AVO, and does NVIDIA use it to power its own AI models? AVO (Agentic Variation Operators) is an agent architecture, not a language model. It was first used to optimize GPU code and was later applied to ARC-AGI-3, in both cases acting as a system wrapped around existing models rather than a model itself.
What’s the difference between ARC-AGI-3’s public and private test sets? ARC-AGI-3 has 135 total environments. Only 25 are public; the remaining 110 are split between a semi-private set and a fully private set, both held back specifically so systems can’t be optimized against known answers.
Which AI model did NVIDIA use inside the AVO harness? The headline 100% result used Anthropic’s Claude Opus 5. NVIDIA separately tested AVO with OpenAI’s GPT-5.6 (“Sol”) on a limited subset of tasks.
Is NVIDIA’s AVO architecture open source or available to developers? No. There’s no public code, API, or SDK for AVO. Available reporting describes it as a closed, internal research system.
How is AVO different from other ARC-AGI-3 systems like VISTA? VISTA runs models through tools like Claude Code or Codex and primarily uses image-based views of each puzzle. AVO uses its own task interface, feeds the model text-based grids instead of images, and adds its own memory and supervision layer. NVIDIA describes its efficiency comparison to VISTA as informal rather than a controlled test.
Does this result mean AI has achieved AGI? No. It’s a strong result on one benchmark’s public test set, achieved with the help of an elaborate support system. It hasn’t been independently verified, doesn’t cover the benchmark’s hidden environments, and researchers who study this area haven’t weighed in on it publicly yet.