A finance team is budgeting for an assistant built on a hosted foundation model, and plans to forecast cost per active user per month. What should they be told?

• ACost is billed per token of input and output, so it scales with conversation length and document size rather than with headcount.
• BCost is billed per session, so a per user forecast is broadly sound.
• CCost is fixed once the model is selected, because inference is included.
• DCost depends on the parameter count of the model rather than on usage.