On "The Pelican comparison grid for Astra is pretty interesting"

Reasoning levels matter way more for visual composition than most people assume, and watching the gap widen across a simple task like this proves it in five minutes instead of five weeks of production work. If you're relying on vision models for anything structural—layouts, SVGs, design logic—you're going to want to test at reasoning=max and accept the latency cost, because the difference between "acceptable" and "actually usable" isn't marginal. This is the kind of quick validation that should be your standard move before committing to any model in a real pipeline.