The main value of regexes is expressing a fairly convoluted ask in a fairly concise form. While I’d expect LLMs to eventually generate regexes correctly, I’m not sure how you’re going to prompt the thing in a way that’s not harder than writing regexes directly.

Excluding the classic problems where you shouldn’t be using a regex in the first place (parsing email, names, html, etc) and the “draw the rest of the owl” where it simply solves everything from the input alone

I think regex is going to be one of the last things LLMs really master, but I'd be happy to be proven wrong.