I imagine they need the compute. Can expand market share with more users for same amount of compute.

But I agree with you. I have a dumb workflow that worked well with v4-flash-0731 and I suspect is directing to a newer model that now breaks it.