I don't know.. It seems stupid easy to me. Try something like "Type the third and fifth letters of the word elephant into the box below" with instagram filter applied. A bunch of variations for the first part like "the following word" and "the word in parenthesis".

Basically, I can't think of a way to come up with something ez that defeats it. You would have to train a neural net specifically on these images because normally neural nets are bad at instagram filter removal unless trained on it. Plus you can slow down/ban/mess with requests based on cookies.

There are basically infinite solutions. Would take a couple of days to implement and would be really fun. I guarantee you: if your site gets maybe 30k visitors a week, nobody would bother spending a month cracking your captcha when there are much easier targets out there.

Finally, you can make it super annoying to actually find where the image is by converting to svg and messing with html structure/compose image in JS. Now they're going to be forced to run a headless browser, take screenshots of the captcha page and finding the image within the page.

If they take the pay-per-captcha approach, I don't think anti-captcha and the like would make it too ez. Still days of work to set up something really brittle.