How much less could this client pay for the same work?

Describe the workloads a client runs today, each with its use case, and read a ladder of savings, largest first: a batch or latency-tolerant tier where nobody is waiting, a cached prompt prefix, a smaller model of the same family or a route to one, another cloud or hosting mode, shorter requests, a cheaper region. Every row is that one change priced from the vendors' list prices at the client's own volumes; the applied changes are priced together as the optimised plan. Nothing is projected and no vendor is favoured.