Nvidia’s AVO is a general-purpose architecture for long-horizon autonomous agents; wrapped around Claude Opus 5 it scored 100.00 on the ARC-AGI-3 public set, completing all 183 levels across 25 environments where the bare model’s baseline was 30% (NVIDIA blog). The week before, the same agent had explored more than 500 directions on DGX B200 systems and produced kernels beating cuDNN by up to 3.5% and FlashAttention-4 by up to 10.5%. The newsletter pairs it with Inherent’s research-taste result as the systems-side proof that architecture, not model capability alone, now moves the frontier. In the trajectory it closes the benchmark that returned frontier models to near zero in March, eight days after stock Claude Code lifted the same model to 96.2%, and joins the 232x kernel and KernelBench results on the line where agents rewrite the kernels they run on.
Development
An agent architecture lifts a model from 30% to a perfect 100
The systems-side twin of the research-taste result: a general-purpose agent architecture rather than a new checkpoint closes the benchmark launched five months earlier to humble the frontier, and the same architecture had just spent a week evolving the kernels it runs on.
Claim
Nvidia's AVO agent architecture lifted Claude Opus 5 from a 30% baseline to a perfect 100 on ARC-AGI-3, sweeping all 183 levels, fresh off a week evolving GPU kernels past FlashAttention-4.
Reported metric: 30% to 100%
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to August 23, 2026 (theinnermostloop.substack.com)
- NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents (developer.nvidia.com)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.