A Token Bill Rewards Careful Reading, Not Guesswork

A per-token bill becomes readable when you treat it as a record of measured work rather than a single price. Each line connects a request, a token category, a rate, and a charge. The practical task is to verify those connections before judging whether the total looks reasonable.

This walkthrough uses invented rates and token counts, not a current price quote. Follow the sequence with one completed request, then repeat it across the billing period.

First, match each charge to its request

Start with the request identifier, timestamp, model name, and completion status. These fields establish the work being billed. Check the billing timezone before comparing dates with application logs, particularly around midnight or the end of a month.

Also distinguish an application action from an API request. One coding task can trigger several requests through tool use, retries, or follow-up reasoning. A repeated timestamp does not by itself prove duplicate billing; compare request identifiers and your client logs.

BestRelay is a quality-first API relay that forwards each request to the official model named in it, without substitution. That makes the requested model an essential reconciliation field rather than a detail to infer from response quality.

Second, separate token categories before applying rates

Read the usage categories before doing arithmetic. Depending on the provider and endpoint, a record may distinguish ordinary input, output, cache creation, and cache reads. BestRelay bills pay-as-you-go from provider-reported token counts, with cache reads and writes priced separately.

Do not assume every displayed input count is an additional billable bucket. Some usage formats include cached tokens within a broader input total; others expose categories differently. Establish whether the counts overlap before multiplying anything. Otherwise, the same tokens can enter your reconstruction twice.

For this example, suppose you have already normalized the record into four nonoverlapping categories: 12,000 ordinary input tokens, 8,000 cache-write tokens, 80,000 cache-read tokens, and 2,000 output tokens.

Third, put every rate on matching units

Assume the example rates are three dollars per million ordinary input tokens, three dollars and seventy-five cents per million cache-write tokens, thirty cents per million cache-read tokens, and fifteen dollars per million output tokens.

For each category, divide its token count by one million, then multiply by its rate. If your actual bill quotes rates per thousand tokens instead, use one thousand as the divisor. Confirm the currency as well. A correct calculation with the wrong unit can still produce a wildly incorrect estimate.

Fourth, calculate each charge before adding totals

The ordinary input charge is 3.6 cents. The cache-write charge is 3 cents. Cache reads contribute 2.4 cents, and output contributes another 3 cents. Together, those four categories cost 12 cents for this hypothetical request.

This breakdown reveals a useful finding: the largest token category need not produce the largest charge. Here, cache reads account for most tokens but cost less than ordinary input. Comparing raw token totals alone hides the effect of different rates.

Keep more precision during calculation than the interface displays. If individual rows are rounded for presentation, adding those visible amounts may differ slightly from a total calculated from unrounded charges.

Fifth, reconcile usage separately from account movements

Now compare the reconstructed request charge with the usage ledger. Then inspect top-ups, credits, refunds, and other adjustments separately. Money added to an account is not token consumption, and a balance change is not necessarily a model charge.

BestRelay members also earn referral commission when people they invite top up. Keep that commission distinct from request costs when reconciling account activity. Create a BestRelay account to inspect your own usage records.

Finally, repeat the check across your tools

BestRelay works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK. Across those clients, use the same reconciliation sequence: identify the request, normalize token categories, verify rate units, calculate charges, and separate usage from account movements.

If a discrepancy remains, preserve the request identifier, provider-reported usage, applicable rates, and your calculation. Those details turn a vague concern about an expensive bill into a specific, testable billing question.