NVIDIA's AVO Hits 100 on ARC-AGI-3, Uses 12% Fewer Actions
TL;DR
- Claude Opus 5 scores 30% on ARC-AGI-3 standalone and 100% inside AVO, a 70-point lift from scaffolding with no model change.
- AVO completed all 183 public-set levels in 6,624 environment actions, 12% fewer than prior leader VISTA's 7,542.
- The 100% result covers only the 25-environment public set; ARC Prize's semi-private and private held-out sets remain untested.
NVIDIA says its new coding agent cleared every level of the ARC-AGI-3 public set.
In a blog post dated August 21, NVIDIA reports that the system, called Agentic Variation Operators (AVO), scored 100.00 RHAE across all 183 levels in the benchmark's 25 environments. On the same task, running the same Claude Opus 5 backbone, VISTA used 7,542 environment actions to AVO's 6,624, a roughly 12% reduction. Claude Opus 5 by itself reportedly reaches 30% at high reasoning effort.
"A frontier language model is only one component of an AI agent," the authors write. "The surrounding agent system… determines how the model receives context, uses tools, maintains state, responds to feedback, recovers from failure, and sustains progress over long-running tasks."
The harness, as the post describes it, has three pieces: a persistent memory that carries implementations and reasoning between attempts, a supervision loop that flags stagnation, and an agent loop that inspects, plans, implements, and evaluates. Similar scaffolding claims have run through our agents coverage all summer.
The team also applied AVO to GPU-kernel optimization, reporting 40 committed kernel versions from more than 500 explored directions over seven days, with the best beating FlashAttention-4 by up to 10.5% on DGX B200 systems. The post is credited to NVIDIA principal engineer Terry Chen along with Jean-Francois Puget and Humphrey Shi, the company's VP of High-Performance AI, among the co-authors.
The framing is blunt. "The model matters, but the model is not the entire agent."
What others are reporting
-
AlphaSignal Read →
Adds the 6,624 vs 7,542 action count comparison, 7-day compute cost, and GPU kernel validation history; frames architecture as the sole driver of the 70-point gap.
The same loop that iterates on CUDA code also iterates on hypotheses about invisible game rules.
-
Crypto Briefing Read →
Stresses AVO is a research demo with no commercialization timeline and explicitly flags the public-set-only limitation as a material caveat.
That gap between 30% and 100% is entirely attributable to the engineering wrapped around the model, not the model itself.
-
Saiyam Pathak (Substack) Read →
Argues agent harnesses now warrant the same versioning and budget scrutiny as models themselves, reframing organizational AI investment calculus.
Claude Opus 5 on its own scores about 30% on the public set. Wrapped in AVO...the same model finishes everything.
-
OfficeChai Read →
Tempers the headline: AVO runs on Nvidia's own reimplementation of the public set only, so the result is a system score tied to specific infrastructure, not a portable model score.
The 100% is best read as evidence that a well-built harness can extract dramatically more from an existing model on this kind of task than the model manages on its own.
Shared on Bluesky by 1 AI expert
Originally reported by developer.nvidia.com
Read the original article →Original headline: Nvidia's AVO Coding Agent Scores Perfect 100 on ARC-AGI-3 Public Set, Beating VISTA With 12% Fewer Actions