Read Token Prices Like a Receipt, Not a Headline

01 / A $1 headline is not a $1 shopping basket

A price list says $1 per million tokens. Your application processes 250,000 tokens. The bill should be $0.25, right? Only if every token falls into that exact $1 category. If input, output, cache reads, and cache writes have different rates, multiplying the total token count by the cheapest visible number gives you the wrong receipt.

The useful question is not “What is the price per million?” It is “Which tokens get which price?” Here is a worked example you can copy into a spreadsheet. Every rate below is illustrative, not a current CheapRelay quote or a claim about any particular model.

02 / Build a four-line receipt for one request

Assume a request has 220,000 input tokens: 120,000 ordinary input tokens, 80,000 cache-read tokens, and 20,000 cache-write tokens. It then generates 30,000 output tokens. That makes 250,000 tokens altogether, but four separately priced buckets.

• ■ Ordinary input: 120,000 tokens at $1.00 per million = $0.120.
• ■ Cache reads: 80,000 tokens at $0.10 per million = $0.008.
• ■ Cache writes: 20,000 tokens at $1.25 per million = $0.025.
• ■ Output: 30,000 tokens at $4.00 per million = $0.120.

Total: $0.273 for the request. The calculation for each line is token count divided by 1,000,000, multiplied by that category’s rate. Add the lines only after pricing them separately.

Notice what dominates: output costs as much as ordinary input despite containing one-quarter as many tokens. Meanwhile, the 80,000 cache-read tokens cost less than one cent. A single blended token count hides both facts.

03 / Check whether cache writes replace or supplement input charges

This receipt assumes mutually exclusive input buckets. The $1.25 cache-write rate is the complete charge for those 20,000 tokens, not an extra fee added to ordinary input pricing. That distinction belongs in your calculation, not in a footnote you discover after topping up.

If a price list instead describes a cache-write surcharge, the arithmetic changes. Charging those same 20,000 tokens at $1.00 ordinary input plus a $1.25 write surcharge would cost $0.045, rather than $0.025. The request total would become $0.293.

Also check the usage fields before adding them. If a reported input total already includes cached tokens, adding that total to cache reads and writes double-counts part of the request. Reconcile the buckets back to the input total first.

04 / Turn $0.273 into a workload budget

At exactly the example’s usage mix, 1,000 requests cost $273. At 30,000 requests, the token subtotal is $8,190. Those are projections from the example, not forecasts of your application. Real requests vary, so keep the assumptions beside the numbers.

Now change just one variable: output grows from 30,000 to 90,000 tokens. Its line rises from $0.120 to $0.360, and the request total becomes $0.513. Across 1,000 requests, that is $513 instead of $273. The price list has not changed; the workload has.

For comparison, pricing all 220,000 input tokens as ordinary input, with no cache-read savings or cache-write premium, gives $0.220 input plus $0.120 output: $0.340 total. Under these illustrative rates, the original cache mix saves $0.067 per request. That saving depends on actually achieving the stated mix.

05 / Ask for the receipt before choosing the route

CheapRelay’s pay-as-you-go setup has no monthly fee or subscription, and billing follows real token usage, including cache reads and writes. That makes the category breakdown especially useful: your budget should follow the tokens your application consumes, not a subscription allowance.

Before comparing routes, put the same input, output, and cache assumptions against each applicable price list. Keep full precision during calculation and round for display afterward; rounding $0.008 to $0.01 before scaling would overstate that line by $2 across 1,000 identical requests.

Sign up for CheapRelay, then use your actual usage breakdown and the applicable rates to build this receipt for your workload. The lowest headline price is only one line; the total is what you pay.