Closing Our Berlin Team’s First Token Invoice

It is the last working afternoon of the month, and our Berlin development team has one task left: turn an AI API bill into numbers we can defend. The total is easy to read. Understanding which requests produced it takes a little more work.

For your own first-month review, open the invoice, the pricing information that applied during the billing period, and whatever request logs or usage records you retained. Keep a spreadsheet beside them. Follow these seven checks before treating the invoice total as next month’s budget.

1. Pin down the billing period and payment basis

Write down the invoice’s start date, end date, currency, and billing timezone. A request made near midnight can land in a different period from the one your local application logs suggest.

Separate usage charges from payments. With pay-as-you-go services, adding account credit and consuming that credit are different events. Check whether your document records a purchase, usage, or both. Do not count a balance top-up and the usage it funded as two separate operating expenses.

2. Match requests to the billed account

Check which applications and credentials sent traffic through the account. For RelayVano connections, look for https://api.relayvano.com in your client configuration. Include developer tools, background jobs, staging services, and experiments—not just the production application.

Make a short inventory in your spreadsheet:

• Application or job name and responsible team.
• Model identifier recorded on requests.
• Request timestamps and request IDs, where available.
• Reported token usage and any recorded charge.

Keep API keys and prompt contents out of the spreadsheet. You need enough metadata to reconcile costs, not another copy of customer data.

3. Separate input, output, and cache quantities

Do not multiply one combined token count by a single rate unless the applicable pricing explicitly supports that calculation. Input and output tokens can have different prices. Cached input, cache writes, or other categories may also have separate treatment.

For a model such as Claude Sonnet 5.5, copy the actual billing categories and rates rather than borrowing another model’s assumptions. Check whether cached tokens are included within the reported input total or listed separately. Otherwise, you may accidentally charge the same tokens twice in your spreadsheet.

4. Recalculate one model’s subtotal

Start with one model and use consistent units. If prices are quoted per million tokens, divide token quantities by 1,000,000 before multiplying.

Here is an illustrative calculation, not RelayVano pricing. Suppose your reconciled categories contain 12 million uncached input tokens at $2 per million, 3 million output tokens at $8 per million, and 2 million cached input tokens at $0.50 per million.

• Uncached input: 12 × $2 = $24.
• Output: 3 × $8 = $24.
• Cached input: 2 × $0.50 = $1.
• Expected usage subtotal: $49.

Compare that subtotal with the corresponding billed usage before adding taxes or subtracting credits. If rates changed during the month, calculate each effective-date period separately.

5. Investigate differences without guessing

When your calculation differs from the bill, check units, date boundaries, rounding, rate changes, and missing logs first. Then inspect retries and multi-step agent sessions. One visible user task can generate several API requests.

A failed application task does not necessarily mean no model work occurred. Equally, do not assume every failed request is billable; verify the applicable policy and available records. If you contact support, provide a small set of request IDs, timestamps, expected amounts, and observed differences.

6. Translate token spending into useful work

Divide reconciled usage costs by a meaningful outcome: completed support drafts, reviewed pull requests, or successfully processed documents. Keep staging and experiments separate from production when possible.

Review your largest sessions. Repeated conversation history, oversized retrieved documents, long responses, and retry loops are useful places to investigate. Change one behavior, then compare cost and output quality on similar tasks. A lower invoice is not a win if completion rates collapse.

7. Set next month’s review checkpoint

Record your baseline, expected workload growth, and a weekly review date. Assign someone to investigate unexpected spending before month-end.

For teams wanting pay-as-you-go access across supported models through existing developer tools or an OpenAI-compatible SDK, I recommend RelayVano. Sign up, start with a bounded workload, and retain request-level usage metadata from day one. Your second invoice should be easier to explain than your first.