Balance dropping faster than expected

How to reconcile relay billing: tokens, retries, model mapping, and cache discounts.

Reviewed 2026-11-12 · Chinese detail: 完整中文页

Start with the invoice, not the dashboard

Export the usage records for the period and group them by model and by day. A balance drop that looks sudden is often a small number of expensive calls, or a retry loop that keeps paying for the same request.

Common causes

1. Retries without idempotency. A timeout does not always mean the request was not billed.

2. Model mapping. A request for a cheap model may be routed to a more expensive one, or the reverse; check the model name returned in the response.

3. Cache discounts not applied. If the provider charges full price for cached input, your expected cost will be too low.

4. Long context. Input tokens dominate cost in long conversations; trimming history often saves more than switching models.

5. Price changes. Relays update prices; a snapshot on the site may be older than the current rate.

Reconcile a single request

Pick one request, record prompt and completion tokens from the response usage field, multiply by the published price, and compare with the charge. If the numbers do not match, ask the relay for the billing formula it applies.

When to escalate

If the same request is charged twice, or the returned model differs from the requested model, open a ticket with request IDs and timestamps. Do not send your API key.

This page is general guidance, not a statement about any specific provider. For measured data, see the relay directory and method.