$0.024 a Request: Read the Whole Token Price

01 / A $1 headline can hide a $0.024 request

A price list says $1 per million tokens. You have a $1 budget. That sounds like plenty, until you discover that the headline covers input only. Generated output has a different rate, and your application keeps sending its growing conversation back with every request.

The useful number is not the cheapest rate on the page. It is the total cost of one useful result. Start there before choosing a model or connecting an automated coding tool.

02 / Turn two rates into one receipt

Here is a made-up price schedule for learning the arithmetic, not a quote from CheapRelay or any provider: input costs $1 per million tokens, and output costs $8 per million tokens.

Suppose one request sends 8,000 input tokens and receives 2,000 output tokens. Its receipt is:

• ■ Input: 8,000 ÷ 1,000,000 × $1 = $0.008.
• ■ Output: 2,000 ÷ 1,000,000 × $8 = $0.016.
• ■ Total: $0.024, or 2.4 cents.

Output represents only one-fifth of the tokens but two-thirds of the bill. At this exact usage, 40 requests cost $0.96, leaving four cents from a $1 budget. That is a much clearer purchasing decision than “$1 per million.”

03 / Count everything sent, not just your latest question

Input can include system instructions, earlier messages, retrieved documents, tool results and attached code. A short question does not guarantee a short input.

For example, sending the same 8,000-token context ten times means processing 80,000 input tokens across those calls. It is not one 8,000-token purchase simply because the text stayed the same. Caching may change the rate, but you need evidence that the requests qualified.

Use the API's reported usage rather than estimating tokens from word count. Before trusting a price comparison, also check whether long-context requests have different rates and whether billable output includes reasoning tokens that are not displayed as ordinary answer text.

04 / Cheap cache reads are not free memory

A cache price deserves its own calculation. Extend the fictional schedule with cache reads at $0.10 per million tokens. If a later request reports 6,000 cached input tokens and 2,000 uncached input tokens, with the same 2,000-token output, the receipt becomes:

• ■ Cache reads: 6,000 ÷ 1,000,000 × $0.10 = $0.0006.
• ■ Uncached input: 2,000 ÷ 1,000,000 × $1 = $0.002.
• ■ Output: 2,000 ÷ 1,000,000 × $8 = $0.016.
• ■ Total: $0.0186, before any applicable cache-write charge.

Do not apply the read rate to the first request automatically. Check cache-write pricing, eligibility and expiration. Also establish whether cached tokens are included within reported input totals or listed separately; counting them twice creates a false bill estimate.

05 / Give a $1 experiment a hard boundary

Start with three to five representative requests, not an unattended overnight run. Record actual usage, actual charges and whether each result was usable. A cheap response that needs three retries can cost more than a successful first attempt.

• ■ Set an output limit appropriate to the task; a requested limit is a ceiling, not a spending target.
• ■ Limit agent steps and retries so one task cannot quietly become twenty calls.
• ■ Remove irrelevant history and oversized tool results before resending them.
• ■ Track cumulative spend locally if the service does not provide an enforceable budget cap.

At CheapRelay, billing is per real token, including cache reads and writes, with no monthly fee or subscription. That suits a small experiment, but pay-as-you-go still needs a stopping rule. Sign up for CheapRelay when you are ready to test a small, measured workload.

06 / Buy useful results, not the lowest table entry

Before adding funds, identify input, output, cache-read and cache-write rates for the exact model you intend to use. Check any applicable pricing conditions. Then calculate cost per accepted result: total experiment spending divided by results you would actually keep. For a tiny budget, that receipt beats a headline rate every time.