You're missing the fact that they went from something that is quite difficult to get a program to do (essentially an AI complete task) to one that's not AI complete.
It doesn't really matter how many non-AI complete components they are measuring... without at least one AI complete task, they removed the thing that makes CAPTCHAs work.
Except distorted text reading is not hard for AI anymore, and all "AI-complete" tasks, whatever that means, are pretty much broken nowadays.
So where left with a very hard problem, and their best solution so far seems to be security through obscurity with a bunch of non-disclosed "cues". Not great, but I guess it's hard to come up with anything better.
It doesn't really matter how many non-AI complete components they are measuring... without at least one AI complete task, they removed the thing that makes CAPTCHAs work.