FM Inference Costs: Token Usage, Provisioned Throughput, and Usage Optimization