An agent built by NVIDIA, called AVO, scored a perfect 100 percent on the public set of ARC-AGI-3, clearing all 183 levels across 25 games. The run was reported by Wccftech on 21 August 2026, which credits the result to the setup around the model rather than to the model itself. On the same public set, Anthropic’s Claude Opus 5, run without that scaffolding, scored 30 percent.
What is AVO and what did it solve?
AVO is NVIDIA’s coding agent. It is a software wrapper around Anthropic’s Claude Opus 5. The benchmark in question is ARC-AGI-3, a public reasoning test designed to measure how systems handle novel, grid-based puzzles. The public set contains 183 levels spread across 25 games.
How did AVO reach a perfect score?
The system received no rules, no prior instruction, and no stated goals for the ARC-AGI-3 tasks. It learned by attempting a level, observing the result, and correcting its next move. Wccftech frames that as the difference between a raw model and an agent: the model handles reasoning, and the surrounding code runs the loop that turns attempts into progress.
To run AVO on ARC-AGI-3, NVIDIA replaced the GPU engineering tools the harness was originally built to use with the ARC-AGI-3 task interface. The underlying agent architecture stayed the same.
What is the 30 percent to 100 percent gap?
Without the loop, Claude Opus 5 scored 30 percent on the same public set. Add the scaffolding and the same model clears every level. The jump is a clean demonstration of how much of an agent’s score comes from the scaffolding rather than the weights.
How efficient was the run?
AVO cleared all 183 levels in 6,624 actions. On the same benchmark, a previous system called VISTA needed 7,542 actions to clear the public set. That works out to AVO being about 12 percent more action-efficient than VISTA.
Where did AVO come from?
AVO was not built for puzzles. It was originally designed to optimize CUDA GPU kernels. In that earlier work it ran autonomously for seven days, explored more than 500 directions, and produced kernels that may beat FlashAttention-4 by up to 10.5 percent. The ARC-AGI-3 run reuses the same agent loop, pointed at a different interface.
What does the result not tell us?
ARC-AGI-3 does not allow external tools to run against its hidden private set. That means AVO’s performance on the private benchmark is unknown. The 100 percent score applies only to the 183 public levels. Any claim about private-set results would not be supported by the public numbers.
Why the harness matters
The ARC-AGI-3 numbers line up with a broader pattern in agent benchmarks: the scaffolding around a model, the tools, the retry loop, and the action budget, often changes a score more than swapping the underlying model does. Same model, different wrapper, 30 percent versus 100 percent.
FAQ
What is AVO?
AVO is NVIDIA’s coding agent. It functions as a software wrapper around Anthropic’s Claude Opus 5.
What did AVO score on ARC-AGI-3?
AVO scored 100 percent on the ARC-AGI-3 public set, clearing all 183 levels across 25 public games. Claude Opus 5 on its own scored 30 percent on the same set.
Does AVO’s score cover the private ARC-AGI-3 set?
No. ARC-AGI-3 does not allow external tools to run against its hidden private set, so AVO’s private-set performance is unknown.
Want help with this in Myrtle Beach or Long Island? Talk to us.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.
We use your details only to reply to you and send what you asked for. See our Privacy Policy.
