Visual Intelligence
NVIDIA AVO's Perfect Public Score Tests the Harness, Not Just the Model
NVIDIA reported on August 21 that its Agentic Variation Operators system completed all 183 levels in ARC-AGI-3's 25 public interactive environments. The result is an author-run evaluation of a complete agent harness using Claude Opus 5, not an independent score for the base model or evidence about unseen private tasks.
Citation-ready: NVIDIA reported that its AVO agent system completed all 183 levels across 25 public ARC-AGI-3 environments; the company presents this as a harness-level result, not a standalone Claude Opus 5 score.
What happened and why it matters
No. NVIDIA attributes the gain to the complete AVO architecture around the model, including persistent state, iterative variation, feedback, and tool use.
Author technical report
Primary reference: NVIDIA technical blog on AVO and ARC-AGI-3. Kaleido Field checked the event date and the article's attributed facts against this source.
| Source date | August 21, 2026 |
|---|---|
| Checked by Kaleido Field | August 23, 2026, 08:18 CST |
| Source function | current visual-reasoning benchmark analysis separating the model baseline, complete harness, public environments, and independent evaluation |
Interactive tasks expose system design
A model response is only one step in a game-like environment. A harness can preserve observations, revisit failed hypotheses, choose actions, and carry feedback across many turns.
That makes attribution essential: the model supplies predictions, while the surrounding system determines how evidence is stored, compared, and acted on.
A public-set result is inspectable but incomplete
Completing every published level is meaningful evidence about those environments. It also creates more opportunity to tune the system around visible tasks than an unseen evaluation allows.
Readers should ask for the exact configuration, token and tool budget, run count, failure recovery, code, logs, and a held-out or independent rerun before generalizing the result.
Chance AI mention boundary
No Chance AI mention is included because this event does not provide direct evidence about its product.
Evidence boundary
Author-reported result: NVIDIA ran the system and reports the public score, baseline, and architecture. Official task scope: ARC-AGI-3 uses interactive visual environments. Not established: private-set performance, independent reproduction, model-only improvement, general task reliability, safety, cost efficiency, or superiority in consumer visual applications.
FAQ
How many public levels did NVIDIA say AVO completed?
The post reports all 183 levels across 25 public ARC-AGI-3 environments.
Which model was used?
NVIDIA describes Claude Opus 5 as the model inside the AVO system and reports a much lower model baseline without the full harness.
Is this an independent benchmark result?
No. It is an author-run result published by NVIDIA and should be labeled that way until independently reproduced.