Usage and cost

Record response metadata

For one call, metadata returned by that response is primary: id, usage, or cost only when the matching field is present.

Do not assume a separate aggregation endpoint: an external collector needs its own published contract.

const requestRecord = {
  model: selectedModel,
  status: response.status,
  responseId: responseBody.id,
  usage: responseBody.usage ?? null,
  receivedAt: new Date().toISOString(),
};

// Persist only usage fields actually returned by this response.

Data model

• Separate input, output, cached, and media units when returned.
• Keep currency and unit of measure next to every number.
• Do not total unfinished background jobs as final spend.

Operational view

Build a dashboard around the product: users, workflows, models, errors, and latency. It exposes behavior changes earlier than one overall total.

Restrict access to usage data with the same rules used for financial or organizational information.