LLM API Budget Basics: Requests, Tokens, and Fixed Costs
Recurring AI spend is volume × average tokens × rates, plus anything fixed (seat fees, vector DB, observability). This guide shows how the LLM API cost worksheet keeps those layers separate.
Variable vs fixed layers
Variable cost tracks metered tokens. Fixed cost covers anything that does not scale with a single request. Mixing them into one “per chat” number hides which lever actually moves the bill.
Build a monthly envelope
Estimate requests
Monthly successful calls, not peak QPS alone.
Attach average tokens
Separate input and output averages from production logs.
Apply your rate card
Paste current provider rates; do not rely on memory.
Add fixed costs
Hosting, evaluation, and support tools that are not token-metered.
Sibling workloads
Fine-tuning or pretraining belongs on the training-cost worksheet. Generative video minutes belong on the video-cost worksheet. Keep token chat budgets distinct so you can compare unit economics honestly.