If you've ever been told that AI is "just autocomplete" — that it doesn't really reason, it just predicts the next word really well — ARC-AGI-3 is the benchmark that was supposed to prove that. It's a test built specifically to separate pattern-matching from actual abstract reasoning: a series of interactive puzzles designed so that humans score 100% and AI systems, as of this past March, were scoring below 1%.

Last week, Nvidia published results showing their system AVO — Agentic Variation Operators — hit 100% on the ARC-AGI-3 public set. All 183 levels. The model powering it: Claude Opus 5, which on its own scores about 30% on the same benchmark.

That 70-point gap is the story. Not the headline.

What ARC-AGI-3 actually tests

ARC-AGI-3 was designed by François Chollet and the ARC Prize Foundation to test something specific: can an AI system acquire new problem-solving strategies for genuinely novel tasks it's never seen before? Not harder pattern matching. Not a bigger training set. Novel interactive reasoning, the kind humans do naturally when dropped into an unfamiliar situation with no instructions.

If you don't build AI: think of it like this. You can teach someone to win at chess by having them memorize millions of games. Drop them into a new boardgame they've never seen and ask them to figure out the rules — the memorization advantage mostly vanishes. ARC-AGI-3 is that second thing. The claim was that AI could do the first thing but not the second. The March 2026 technical report showed frontier models below 1% on their own, while humans hit 100%.

What AVO does

AVO is not a model. It's a scaffolding system that wraps Claude Opus 5 and gives it:

• Persistent memory across the full task — it doesn't forget what it tried an hour ago
• A supervisor that watches the agent and re-routes when it gets stuck
• An iterative loop that keeps trying, inspects failures, and adjusts strategy
• Swappable tools — write code, run commands, check documentation, validate via execution

The result: AVO completed all 183 levels in 6,624 environment actions, 12% fewer than the prior top scorer's 7,542. Real efficiency, not brute force.

What the fine print says

Nvidia is explicit about the limits:

Public set only. The 100% result covers 25 environments and 183 levels. ARC Prize's semi-private and private held-out sets — the ones that matter for a real benchmark claim — remain untested. You cannot read "100% on the public set" as evidence of a generalizable breakthrough.

Not a controlled ablation. The 70-point gap between Claude Opus 5 alone and AVO-wrapped Opus 5 cannot be cleanly attributed to AVO's architecture. The two test conditions differ in more than scaffolding: different agent backends, different observation representations, different memory systems, different reasoning-effort settings. Nvidia says so directly in their writeup. You can't subtract 30 from 100 and call the result "what AVO adds."

The underlying model didn't change. Claude Opus 5 is still Claude Opus 5. What changed is the system around it. That distinction matters enormously for what this result means about AI capability.

Source spread

• Nvidia Technical Blog, Aug 21 — hype. The primary source; accurate on the numbers but leads with "100%" without foregrounding the public-set scope.
• ARC Prize Foundation — ARC-AGI-3 Technical Report — academic. Original benchmark design, humans-at-100% / AI-below-1% baseline.
• AlphaSignal — skeptic. Leads with "public set only." Correct framing.
• Forbes, Aug 24 — hype. "Perfect score" headline, caveats inside.
• The New Stack — builder. Focuses on architectural takeaways; more balanced than most.

What's real:

• The 100% public-set result is real. AVO genuinely solved all 183 levels — it wasn't achieved by gaming the environment or exploiting an evaluation loophole.
• The scaffolding architecture is worth studying. Persistent memory across long-horizon tasks, with a supervisor and iterative loop, is the pattern serious agentic systems will converge on. Nvidia published a concrete instantiation of what that looks like.
• The 70-point gap shows something important: what you wrap around a model matters more than a marginal improvement in the model itself. This is the real headline.

What deserves a side-eye:

• The private held-out set is what matters. ARC Prize has always distinguished public from private scores. Acing the practice set and calling it the exam is not how this benchmark works.
• The comparison between 30% and 100% is not a controlled measurement of AVO's contribution. Too many variables changed. Nvidia says so. Most of the coverage ignores that.
• "100%" will land in a thousand posts without the caveats. The framing in Nvidia's own announcement made that easier than it should have been.

What to do about it

• Don't update your priors on AGI yet. This is a meaningful scaffolding result, not a finished argument about AI capability. The harder private-set benchmark remains.
• Do notice the 70-point gap. If you use AI tools for complex work — research, planning, writing, coding — the tool matters less than how you use it. The right structure around a model can change your results as dramatically as switching to a stronger model.
• Watch for private-set scores. If AVO or a successor posts a high score on ARC-AGI-3's private held-out environments, that's the headline worth updating on.
• The "just autocomplete" dismissal is getting harder. Whether this specific result proves it or not, the trajectory is clearly toward AI systems that can reason in novel environments. Plan accordingly.

Further reading

• Nvidia Technical Blog — AVO reaches 100% on ARC-AGI-3 — full architecture write-up from the source
• ARC-AGI-3 Technical Report (arXiv 2603.24621) — benchmark design and the humans-vs-AI baseline
• ARC Prize — What is ARC-AGI? — plain-language overview
• AlphaSignal — AVO hits 100%, public set only — correct framing
• Forbes — Nvidia AVO pushes Claude Opus 5 to a perfect score — Aug 24 coverage