$21.30 or $17.04? Make the Token Receipt Match

01 / A $4.26 difference needs an itemized receipt

An API quote is not a bill. A low input-token rate can look attractive while output tokens, cache writes, or payment fees change the total. The fair comparison is simple: price the same workload, with the same model version and billing categories, through both routes.

The worked receipt below uses hypothetical rates, not verified official list prices or current CheapRelay prices. It demonstrates the arithmetic without presenting an invented promotion as a real offer. Before spending money, replace every rate with a dated quote from the relevant pricing page or account dashboard.

02 / Fix the workload before comparing rates

Start with one representative batch of requests. Record actual usage rather than estimating tokens from character counts. For this example, the batch contains four mutually exclusive billing buckets:

• ■ Ordinary input: 2,000,000 tokens.
• ■ Cache reads: 6,000,000 tokens.
• ■ Cache writes: 400,000 tokens.
• ■ Output: 800,000 tokens.

That is 9.2 million billed tokens, but multiplying that total by one advertised rate would be wrong. Each category has its own price. Also check whether a provider reports cache-write tokens separately or inside a broader input count. Do not charge the same tokens twice when translating usage reports into a spreadsheet.

03 / Receipt A: a hypothetical $21.30 list-price bill

Assume the official list-price column in your worksheet contains the following example rates, all in dollars per million tokens:

• ■ Ordinary input: 2.0 million × $3.00 = $6.00.
• ■ Cache reads: 6.0 million × $0.30 = $1.80.
• ■ Cache writes: 0.4 million × $3.75 = $1.50.
• ■ Output: 0.8 million × $15.00 = $12.00.
• ■ Total token charge: $21.30.

Output accounts for $12.00 of this receipt, more than half the total. That makes output pricing more important here than a headline input rate. A workload dominated by repeated cached prompts could have a different cost driver.

04 / Receipt B: a hypothetical $17.04 relay bill

Now hold the workload constant and apply a second set of illustrative rates. These are worksheet inputs, not a CheapRelay quote:

• ■ Ordinary input: 2.0 million × $2.40 = $4.80.
• ■ Cache reads: 6.0 million × $0.24 = $1.44.
• ■ Cache writes: 0.4 million × $3.00 = $1.20.
• ■ Output: 0.8 million × $12.00 = $9.60.
• ■ Total token charge: $17.04.

The difference is $21.30 − $17.04 = $4.26. Dividing $4.26 by the $21.30 baseline gives 20%. That percentage belongs only to this constructed example; it is not a claimed live discount. With different rates or token proportions, calculate the result again.

05 / Keep list-price savings separate from cash savings

Official list price is a useful public baseline, but it may not be your actual alternative. If your direct account has negotiated pricing, batch rates, or credits, create a third column for the amount you would really pay. Label comparisons accurately: savings against list price and savings against your current bill are different claims.

Keep taxes, currency conversion, payment charges, and unused prepaid balances visible as separate lines where applicable. Do not assume either route has those charges. Check them. Likewise, retries can increase spend: compare completed tasks as well as token totals if the two routes consume different amounts.

06 / Use CheapRelay's token bill as the comparison input

CheapRelay is a low-priced, pay-as-you-go AI API relay with no monthly fee or subscription. It bills per real token, including cache reads and writes. Those categories belong in the worksheet, not in a footnote. Its OpenAI-compatible routes use https://api.cheapapirelay.com/v1; Gemini native routes use https://api.cheapapirelay.com/v1beta.

Sign up for CheapRelay, check the current rates for your chosen model, and run a small representative batch before moving a larger workload. Keep the usage record and resulting charge. The defensible saving is the difference between two matched receipts—not the biggest percentage on a pricing banner.