Update: based on info from design arena, it looks like gpt-image-2.6 was post-trained to output less text: https://x.com/DesignArena/status/2118124381155303918

This may be why text elements in particular appear more "zoomed in", and why it's providing less detail for UI design.