Thanks! It's using Davinci and maxing out the token limit so it's about 8 cents per generation.

Not sure if I understand your second question. GPT-3 is outputting the props of the component, the primitive elements in the component (images, icons, text, badges), the "values" (i.e. which copy should this text field have) the primitive elements should have in terms of the props, and some sample data to set the props to.

A separate layout optimization is determining how to arrange the primitive elements.

Given all that it's possible to generate code for a react component that you can plug whatever values you want into for the props.