Monday Invoice Review at Our Berlin API Team

Fare watch EN2026-10-28

What the invoice is measuring

Your first per-token invoice can look more complicated than it is. At its core, it records how much text your application sent to a model, how much text the model returned, and the price assigned to those units. “Token” means a small piece of text. A word, part of a word, punctuation mark, or space may each count as one or more tokens.

The invoice is not charging for the number of chat messages alone. A short request may include a long system prompt, conversation history, tool instructions, or retrieved documents. Those hidden or repeated parts can make the billed token count much higher than the text you typed.

The date range and billing status

Start at the top of the invoice. Confirm the service period, invoice date, currency, and status. The service period tells you which API activity is included. The invoice date tells you when the bill was created. If you are reviewing costs for a launch or a monthly budget, use the service period rather than the invoice date.

Also check whether the amount is open, paid, pending, or covered by an account credit. A zero balance does not necessarily mean that no usage occurred; it may mean that a credit or adjustment covered the charges.

Input tokens and output tokens on each line

Most per-token invoices separate input and output usage. Input tokens are the material your application sends: the prompt, system message, conversation history, files converted to text, and tool context. Output tokens are the model’s response. Output tokens often have a different rate because generating text requires a different amount of processing.

You may also see cached input, batch usage, or another specialized category. Treat each category as a separate meter. Do not combine them until you have checked the rate shown for each one.

Reading a model line from left to right

A typical line can be read in this order:

Model or service name: the model that handled the request.
Usage quantity: the number of input, output, or other billable tokens.
Unit rate: the price for a defined amount, commonly one million tokens.
Line total: the charge produced by that quantity and rate.

For a simple example, 250,000 input tokens at a rate of 3 dollars per million cost 0.75 dollars. If the same workload produced 50,000 output tokens at 15 dollars per million, the output cost would also be 0.75 dollars. The combined line cost would be 1.50 dollars before any credits, taxes, or adjustments. Use the rates printed on your invoice, not a rate copied from an old estimate.

Why your dashboard and invoice may differ

A dashboard can show live or rounded usage, while an invoice may include finalized totals. Small differences can come from rounding, delayed requests, retries, failed requests that still consumed tokens, or activity recorded near the start or end of the billing period.

Model choice can also change the result quickly. A test using Claude Sonnet 5.5 may have a different input and output rate from one using Gemini 3.8 Flash. Compare both the token count and the rate before deciding that a usage increase is an error.

Finding the requests behind an unexpected total

When a line looks high, narrow the investigation in this order:

Group usage by model and compare it with your deployment configuration.
Check whether a long conversation or system prompt is sent on every request.
Look for retries caused by timeouts, client errors, or automatic fallback logic.
Review tool calls, document retrieval, and large pasted outputs.
Compare the invoice period with your application logs and deployment events.

For a reliable comparison, log a request identifier, model, approximate input size, output size, and timestamp. Avoid logging private prompts unless your security policy allows it.

Using RelayVano data for the next month

RelayVano is a practical pay-as-you-go relay for teams that want one API path across supported AI models. You can use the OpenAI-compatible API base at https://api.relayvano.com with Claude Code, Codex, Gemini CLI, or another compatible SDK. Set a monthly spending limit, review the first few days of usage, and investigate repeated context before it becomes a recurring cost.

After confirming your payment method and intended model, sign up for RelayVano and make one small test request before moving production traffic. That test gives you a known token count and a baseline line item for future invoice reviews.