$0.10 for 10 Requests: Count Tokens First

01 / A huge context window is capacity, not a bill

If your AI budget is tiny, start with the request receipt. A model advertised with a large context window does not automatically charge you for filling that window. Your cost depends on the tokens processed and the applicable rates.

Think of the context window as workspace. It limits how much material the model can handle in a request. For many models, input and generated output share that limit, while output also has a separate maximum. Check the model's rules before assuming the entire advertised window is available for your documents.

Your first comparison should be simple: how many tokens does your actual task need, and does the model have enough room?

02 / Count the whole request, including repeated history

Suppose you ask an assistant to summarize a short policy document. The visible question might be only one sentence, but the API request could contain:

• ■ 500 tokens of instructions describing the task and answer format.
• ■ 4,000 tokens from the policy document.
• ■ 1,000 tokens of earlier conversation.
• ■ 500 tokens in the latest question and supporting notes.

That adds up to 6,000 input tokens. If you allow a 1,000-token answer, you need room for roughly 7,000 tokens under a shared input-and-output limit. Leave additional room when the model's accounting includes other token categories.

These are illustrative counts, not measurements of a particular document. Tokens are pieces of text, and their relationship to words varies by language, formatting, and tokenizer. Use a compatible token counter for planning, then inspect the API's reported usage after a real request.

03 / A hypothetical receipt totals $0.10 for ten calls

To practice the arithmetic, assume an input price of $1 per million tokens and an output price of $4 per million tokens. These are example rates, not a CheapRelay price quote.

• ■ Input: 6,000 ÷ 1,000,000 × $1 = $0.006.
• ■ Output: 1,000 ÷ 1,000,000 × $4 = $0.004.
• ■ Total: $0.010 per request, or $0.10 for ten identical requests.

This example assumes ordinary input and output billing, with no cache charges or other billable categories. Substitute the actual rates and usage for your chosen model. An output limit is a ceiling; the bill depends on the answer actually generated and any other applicable usage.

04 / Compare a 32,000-token window with your 7,000-token task

Imagine two options advertise context windows of 32,000 and 128,000 tokens. Your estimated 7,000-token task fits within either, assuming their output limits also permit the answer. The larger window alone gives you no reason to pay a higher rate.

Test whether the cheaper suitable option produces an acceptable answer. Capacity and answer quality are separate questions: fitting the document does not guarantee that the model will interpret every detail correctly.

If a later request contains 40,000 input tokens, the smaller window becomes a real constraint. You could choose more capacity, extract the relevant sections, or split the work. Splitting may repeat instructions and require a final synthesis, so calculate the total cost across all calls.

05 / Trim the text that gets billed again

Conversation history is an easy expense to miss. If your application resends the policy and every previous message on each turn, later requests can grow even when your new question is short.

Keep necessary facts, remove unrelated exchanges, and request a shorter answer when that serves the task. Where caching is supported, check eligibility and pricing: cache reads and writes can carry charges, so repeated text should never be assumed free.

06 / Use a small trial to set your spending limit

CheapRelay offers pay-as-you-go access with no monthly fee or subscription, billing per real token, including cache reads and writes. Sign up for CheapRelay and begin with a small trial budget using the current rates for your chosen model.

Record input tokens, output tokens, cache usage, and cost for each trial request. Compare those receipts with answer quality before increasing usage. Your own requests provide the most useful evidence for deciding how much context you need.