The interesting aspect of the ongoing tests I feel is seeing how models can plan out an image directly using SVG primatives solely through reasoning (code-to-code). If they have a reference then it's a different type of challenge (optimizing for a trace).