Context Windows: Read the Token Receipt

01 / A large window is capacity, not your receipt

A context-window headline answers a capacity question: how much can fit into one request? It does not tell you how many tokens your application actually uses. Input, output, and potentially reasoning tokens must fit within the applicable limits, so comparing the window against just your latest message misses part of the workload. ([developers.openai.com](https://developers.openai.com/api/docs/guides/conversation-state?utm_source=openai))

The useful comparison starts with an itemized request. Below, every price is an invented accounting assumption, not a CheapRelay quote or a claim about any particular model. The arithmetic shows what to measure before choosing a larger window.

02 / Receipt one: 10,000 input tokens, not 128,000

Assume ordinary input costs $2 per million tokens and output costs $8 per million. Our example request contains the following measured token counts:

• ■ System instructions: 1,000 input tokens.
• ■ Conversation history: 3,000 input tokens.
• ■ Retrieved documents: 5,000 input tokens.
• ■ Current question: 1,000 input tokens.
• ■ Generated answer: 2,000 output tokens.

The input subtotal is 10,000 tokens. Its charge is 10,000 ÷ 1,000,000 × $2 = $0.020. The answer adds 2,000 ÷ 1,000,000 × $8 = $0.016. Total: $0.036 per request, or $360 for 10,000 identical requests.

For this simplified workload, with no additional reasoning tokens, the combined footprint is 12,000 tokens. That occupies 37.5% of a hypothetical 32,000-token window, approximately 9.4% of a 128,000-token window, or 1.2% of a million-token window. Those percentages describe capacity utilization, not answer quality.

03 / Receipt two: the same question with extra baggage

Now attach another 6,000 tokens of documents while holding the answer length and example rates constant. Input rises to 16,000 tokens and costs $0.032. Output still costs $0.016. The new total is $0.048.

That extra attachment adds $0.012 per request, or $120 across 10,000 requests. Nothing in this calculation changes because the available window is larger. The additional charge comes from additional billable tokens.

Before deleting context, ask whether it supports the answer. For this example, removing a duplicate 6,000-token attachment restores the original receipt. Removing the only relevant evidence might lower the bill while defeating the task. Treat trimming as an application test, not an automatic optimization.

04 / Receipt three: separate cache reads from cache writes

Cached input is a separate accounting category, not a synonym for free input. Official token guidance distinguishes ordinary input, cached input, and output; actual rates depend on the selected model and category. ([help.openai.com](https://help.openai.com/en/articles/4936856-understanding-and-counting-tokens?utm_source=openai))

Extend our fictional rate card: cache reads cost $0.20 per million tokens, and cache writes cost $2.50 per million. Assume the write rate replaces the ordinary input charge for that portion; these are mutually exclusive categories in this example. An 8,000-token prefix is written once, then reused nine times:

• ■ First request: $0.020 cache write + $0.004 fresh input + $0.016 output = $0.040.
• ■ Each repeat: $0.0016 cache read + $0.004 fresh input + $0.016 output = $0.0216.
• ■ Ten requests: $0.040 + nine × $0.0216 = $0.2344.
• ■ Ten fully uncached requests: ten × $0.036 = $0.3600.

The example difference is $0.1256. Do not assume those nine reads occurred: inspect reported cache usage and the applicable billing rules. Provider documentation exposes separate cache usage fields precisely because a repeated-looking request is not sufficient evidence of the billed categories. ([developers.openai.com](https://developers.openai.com/api/docs/guides/prompt-caching?utm_source=openai))

05 / Choose capacity after measuring your real payload

Build a small ledger containing ordinary input, cache reads, cache writes, output, and total charge. Keep output visible: prompt caching does not change how output tokens are generated. ([developers.openai.com](https://developers.openai.com/api/docs/guides/prompt-caching?utm_source=openai))

Use actual token counts rather than assuming equal-length text produces equal bills across models. Tokenization and generated output can differ. ([help.openai.com](https://help.openai.com/en/articles/4936856-understanding-and-counting-tokens?utm_source=openai))

CheapRelay fits this receipt-first approach with pay-as-you-go billing for real tokens, including cache reads and writes, and no monthly fee or subscription. Sign up for CheapRelay, check the selected model’s current rates, and price your own representative requests. Buy enough context capacity for the job; judge the expense from the receipt.