NVIDIA's AVO loop takes Opus 5 from 30% to 100% on ARC-AGI-3
The model alone managed 30%; the agent loop around it cleared all 183 levels — and the recipe is a paper you can copy this weekend.
Copy markdown
The wrapper cleared it, the model didn't
Claude Opus 5 on its own scores about 30% on ARC-AGI-3. Wrapped in NVIDIA's AVO agent, the same model hit a perfect 100% — all 183 levels across 25 environments — while using ~12% fewer actions than the prior VISTA system (6,624 vs 7,542).
Steal the loop for your own agent
There's no code or weights yet — just NVIDIA's writeup and an arXiv paper — but the recipe is copyable: a hypothesis-act-observe-revise loop with persistent memory and stagnation detection, wrapped around whatever frontier model you already call. The takeaway for builders: scaffolding, not the base model, is where long-horizon wins hide.
Elsewhere: transparent PNGs straight from gpt-image-2
OpenAI's image API now supports transparent backgrounds in preview for gpt-image-2 — set background to "transparent" and export PNG or WebP (JPEG won't work). Logos, icons, and game sprites with no manual cutout step.