AI-assisted source summary. Read the scope and exceptions, then follow the original documents below.
Applies to
Token-billed AI APIs · documented OpenRouter fields
Time context
Documentation checked October 20, 2026
Sources checked
A billing discrepancy needs investigation. Hidden reasoning or different tokenizers can explain some differences without proving inflated usage.
Ask for an itemized record
OpenRouter's usage documentation exposes native input/output token counts, cost, and available cache and reasoning details. Our suggested comparison is cash paid per completed workload, with input, output, cache rates and the reseller's credit conversion recorded separately.
Reasoning tokens may be charged even when the reasoning text is excluded. OpenRouter notes that reasoning can consume the output budget before a visible answer appears. Counting characters on screen therefore cannot reconstruct the invoice.
Our checklist: save a small non-sensitive test request, request ID, usage, posted rates and balance change. Check whether reasoning is already included in completion totals before adding it again. Ask the seller to explain unexplained multipliers; a seller-controlled usage field alone is not an independent audit.
Original documents used for this summary. Multiple links from one publisher are not independent corroboration. Policies can change after the check date.
AI-assisted research, writing and translation by No Secrets, using the linked public documents. This is a source summary, not firsthand reporting or an independently audited investigation. No affiliate links are used in this article.
The date above records this research pass, not the date of every underlying event. The English and Chinese versions are AI-assisted summaries; the linked original controls if they differ.