$0.0304 for the Request, Not the 128K Window

01 / A 128,000-token window is capacity, not a bill

A large context window tells you how much material a model can work with. It does not tell you how much you send, how much comes back, or what the request costs. Buying API access is not like buying a storage plan: an available 128,000-token window does not automatically become 128,000 billed input tokens.

The useful comparison is therefore not just window size against window size. It is capacity against your actual request receipt. A model with room for an entire repository may still receive only a short instruction, two files and a few messages.

All rates below are hypothetical calculation examples, not CheapRelay prices or quotes for any particular model. Check current rates and the model's context and output limits before budgeting.

02 / The first receipt: 12,000 input tokens and 800 output tokens

Assume an input rate of $2 per million tokens and an output rate of $8 per million tokens. Your request contains instructions, conversation history and selected source files totaling 12,000 input tokens. The response contains 800 output tokens.

• ■ Input: 12,000 ÷ 1,000,000 × $2 = $0.0240.
• ■ Output: 800 ÷ 1,000,000 × $8 = $0.0064.
• ■ Total request cost: $0.0304.
• ■ 1,000 identical requests: $30.40.

For an illustrative 128,000-token context capacity, that input occupies about 9.4% of the advertised window. The unused space does not appear as another input charge. Remember that context and output limits interact differently across APIs; do not assume the entire advertised window is available for input plus unlimited generation.

03 / Filling the window changes the receipt to $0.2624

Now compare that request with one containing 128,000 input tokens at the same illustrative rates, still producing 800 output tokens. This is a billing comparison, not a claim that a particular 128,000-token model accepts that input and output combination.

• ■ Input: 128,000 ÷ 1,000,000 × $2 = $0.2560.
• ■ Output: 800 ÷ 1,000,000 × $8 = $0.0064.
• ■ Total request cost: $0.2624.
• ■ 1,000 identical requests: $262.40.
• ■ Difference from the smaller request: $232.00 per 1,000 requests.

That difference buys additional input processing, not automatically a better answer. Include the extra material when it is relevant. Do not paste an entire project merely because the window can accommodate it.

04 / Five conversation turns can rebill the same history

In a workflow that resends conversation history, each request has its own input receipt. Suppose five requests contain 12,000, 24,000, 36,000, 48,000 and 60,000 input tokens. Together, they bill 180,000 input tokens, not just the final 60,000. Assume each response contains 800 tokens and no cache discounts apply.

• ■ Growing history: 180,000 input tokens cost $0.3600.
• ■ Five responses: 4,000 output tokens cost $0.0320.
• ■ Session total: $0.3920.
• ■ Keeping each request at 12,000 input tokens: $0.1520 total.
• ■ Difference: $0.2400 per session, or $240 across 1,000 sessions.

The smaller session is not guaranteed to preserve answer quality. Trimming history can remove essential decisions. Keep requirements, current errors and relevant code; remove duplicated logs and superseded drafts first.

05 / Cache lines need their own rates

Repeated input may qualify for caching, but repeated text is not proof of a cache hit. Separate uncached input, cache reads, cache writes and output in your cost estimate. Apply the provider's applicable billing rules rather than charging every category at the ordinary input rate.

For example, if 10,000 tokens are billed as cache reads at an illustrative $0.50 per million, that line costs $0.0050. Another 2,000 uncached tokens at $2 per million cost $0.0040. Add the same $0.0064 output charge: $0.0154 total, assuming no additional cache-write charge applies.

06 / Compare three receipts before choosing more capacity

Record a typical request, a long conversation and your largest necessary task. Compare their token usage and costs, then choose enough capacity for the work. CheapRelay provides pay-as-you-go access with no monthly fee or subscription, billing real tokens, including cache reads and writes.

Sign up for CheapRelay to test your own requests, then use reported token usage and current rates to build receipts like these. Budget for what you actually send and receive, not the largest number on a model card.