Each task the agent runs sends text and documents to the model and gets an answer back, and the provider bills by volume. So the monthly cost depends on two things: how many tasks there are each month, and how large each document is.
Simple tasks, like sorting an email or extracting lines from an invoice, can run on small, inexpensive models. Complex tasks, like reading a drawing, get a stronger model. The model is chosen per task for the best price-performance.
The charges go straight to your own account with the provider, with no mark-up from us, and you see every charge. Discovery estimates the expected usage, so there are no surprises. With a local model on your own servers there's no usage billing; the cost is hardware instead.
Choosing a model by cost per task