I have been digging without being able to wrap my head around this. It seems like once a role is deployed, you are charged for it in full, whether or not you scale it up or down?
Why would anyone scale down with this? I don't see the incentive to not just leave the role with all possible instances used to max?
I can see why an availability set, with several roles, might want to distribute the cores between them depending on load. But is that all it's for?
You pay the price of one instance of choosen size (A0 to D14) multiplied per the number of instance that are running.
Try with the Azure Princing Calculator, the number of instance increases charges.
When you try to use Autoscaling it clearly states :
Autoscale is saving you XX% off your current bill