Possible failure modes with this approach: quite a few candidates who are not strong at actually doing the work will be able to somewhat convincingly "talk the talk" well enough to pass stage 2. If stage 3 is done async then there's a minority of candidates who will cheat and get someone else to solve the challenge for them. If the org has set up a hiring pipeline rather that first tries to identify strong hires before figuring out exactly which team they would join rather hiring to fill a specific role then step 2 might not make sense either, as there's no specific team to interview.
I agree that the in person time pressured performative problem solving & coding engineering interview process is often a poor approximation of the work.