Chapter 20 made cost a runtime constraint and showed where the spend comes from. Turning "the gateway processed 1.3 million tokens" into "this customer owes $X, and we earned $Y of margin on it" is a different system with different requirements. Telemetry can drop a span and no one is harmed; a billing pipeline that double-counts or loses events either erodes customer trust or quietly erodes margin. The discipline is to separate the high-volume, real-time metering path from the low-volume, financial-grade invoicing path — and to know exactly where the seams are so you can decide what to build and what to buy.

25.1 The decomposition

Usage becomes money through a chain that most tools blur together; naming the stages tells you where your platform ends and a billing vendor begins. Meter — count what happened. Mediate — deduplicate and normalize the raw events. Rate — apply pricing to produce charges. Bill — aggregate charges over a period. Invoice — render the bill. Collect — take payment. Recognize — book the revenue. The first three are high-volume and real-time; the last four are financial and audited. The boundary between them is the most important architectural line in the whole surface.

25.2 The emission surface — where usage is born

The model gateway is the canonical meter point. It already sees the model, the input/output/cached token split, latency, and success or failure (Ch. 6, Ch. 18), so metering rides on instrumentation you built anyway. Other dimensions join from their own sources: sandbox-seconds (Ch. 9, Ch. 24), tool calls (Ch. 8), retrieval and storage (Ch. 10–11), and seats. Every usage event carries an idempotency key — naturally run_id plus step_id — so that at-least-once delivery deduplicates to effectively-once. And you reconcile metered tokens against the bills you receive from Bedrock, Anthropic, or OpenAI: the gap between what you metered and what you were charged is your margin alarm.

25.3 The stack and the integration strategy

The strategy that holds these together: ingest events idempotently and tolerate late or out-of-order arrival; pre-aggregate so a real-time balance is always queryable; choose prepaid wallets with real-time decrement or postpaid metered invoicing per product; and price across multiple dimensions — per-token (input, output, and cached differ), per-run, per-tool-call, per-seat, a platform fee, a markup on raw model cost. The defining move for an agentic platform is to close the loop: the aggregated balance feeds back into the gateway so an over-budget tenant or runaway agent is throttled or blocked in real time. Metering and the budget enforcement of Chapter 20 are two ends of one mechanism.