Keep three records separate
| Record | Owner | What it answers |
|---|---|---|
| Rate card | Buyer or authorised technical owner | What did the provider officially publish, for which service and date? |
| Usage estimate | Product and technical owners | Which measured or assumed inputs drive cost? |
| Invoice reconciliation | Finance or billing owner | What did the provider report or invoice, and what remains unexplained? |
Do not use a rate card as a workload forecast. Do not call an estimate an invoice. Do not assume a provider's usage dimension maps directly to another provider's token or request count.
Record the actual billable dimensions
Official provider material illustrates why a single tokens-per-month field is too coarse. Current OpenAI pricing and cost-export guidance distinguish pricing and cost reporting. Google's Gemini billing documentation describes input, output, cached-token, and storage dimensions. Anthropic publishes model and usage pricing concepts.
The exact dimensions, tiers, regions, tools, and discounts must be checked on the actual provider page before the buyer relies on them. This article does not state a price, provider ranking, or representative workload.
The blank evidence worksheet keeps input, cached input, output, tools or media, retries, fixed fees, cap, alert threshold, billing owner, reconciliation, and rate-change action visible.
Turn a quote into a control loop
For each metered service, decide before launch:
- who owns the rate card and refresh date;
- which application signals will measure usage;
- the period budget cap and alert threshold;
- who receives the alert and what action they may take;
- who can approve a higher cap or material model change; and
- how the provider report will be compared with the invoice.
A cap without an owner is a notification. A budget without an action is a number. The action might pause a non-critical feature, lower an approved service level, obtain approval, or investigate unexpected retries.
Refresh before each material decision
Use the evidence packet's rate-card refresh procedure immediately before signing a quote, enabling a new model or tool, changing the workload, or approving a cap increase. Record the official URL, access date, currency, dimension, tier, and any fixed or tool charge.
Keep provider workloads separate when comparing options. Equivalent prompts can produce different measured input, output, cache, tool, and retry profiles; a unit-price comparison alone cannot answer the application-cost question.
Reconcile rather than silently revise
At month end, compare the provider's cost or usage export with the invoice and your application telemetry. Retain unexplained differences as exceptions. They may come from timing, tiers, billing configuration, retries, tools, discounts, or a measurement error. Do not change the estimate retrospectively just to make it match.
OpenAI's current guidance describes exporting detailed usage or cost data from its dashboard. That is a provider-specific reporting path, not proof that any estimate will reconcile exactly or that another provider reports the same way.
Acceptance rule
Approve metered AI scope only when the buyer can inspect a dated rate card, workload assumptions, named billing owner, cap, alert action, reconciliation path, and approval rule for rate or scope changes. If any field is unknown, record it as a risk or defer the commitment; do not hide it in an implementation price.