$0.10 Test Budget: Kimi K3 and Claude Sonnet 5

01 / Start with a $0.10 comparison, not a whole book

If you have a tiny budget, do not begin by uploading a 300-page document to both models. Start with a $0.10 spending target and a representative excerpt. Your first question is practical: can Kimi K3 or Claude Sonnet 5 extract the information you need at an acceptable total cost?

There is no honest universal price per page. APIs charge for tokens, while pages vary with layout, language and text density. The numbers below are planning examples, not advertised model prices. Check the current input, output and cache rates for each model before sending a request.

02 / Turn 100 pages into a usable token estimate

For a text-heavy English document, use 500 words per page and 750 tokens per page as a rough worksheet assumption. That makes 100 pages approximately 75,000 input tokens. This is not a measurement: tables, code, unusual formatting and OCR errors can change the count substantially.

• ■ 10 pages: approximately 7,500 document tokens.
• ■ 50 pages: approximately 37,500 document tokens.
• ■ 100 pages: approximately 75,000 document tokens.

Add your instructions and any conversation history to the input estimate. Extract text from the file first when possible, then inspect it for repeated headers and broken OCR. If you submit page images instead, do not assume this text-only calculation still applies. Also verify the selected model's context limit before sending a large file.

03 / A fictional 100-page receipt comes to $0.16

The basic calculation is: input tokens divided by 1,000,000, multiplied by the input rate, plus output tokens divided by 1,000,000, multiplied by the output rate. Rates here mean dollars per million tokens.

To demonstrate the arithmetic only, suppose a model costs $2 per million input tokens and $10 per million output tokens. These are fictional worksheet rates, not a quote for either model.

• ■ 75,000 input tokens cost $0.15.
• ■ 1,000 output tokens cost $0.01.
• ■ Total: $0.16 for 100 pages, or $0.0016 per source page.

That receipt covers one request, not an ongoing conversation. Sending the entire document again can create another large input charge. A long answer also costs more than a short extraction. Divide the complete request bill by source pages processed, rather than counting only the document's input charge.

04 / Compare Kimi K3 and Claude Sonnet 5 on ten pages

Choose ten pages containing a normal section, a difficult table and at least one important exception. Give both models identical extracted text and instructions. Request five findings, supporting page references and a short list of uncertainties. Keep the output allowance small, perhaps 300 tokens, while allowing enough space for useful evidence.

At the fictional rates above, 7,500 input tokens and 300 output tokens cost $0.018 per request before instruction overhead. Two such requests would cost about $0.036. Actual model rates may be higher or lower, so calculate both estimates before treating $0.10 as sufficient. A budget target is not an automatic spending cap.

Score each answer against the source: correct findings, accurate references and missed exceptions. Do not declare a winner from fluent prose alone. The cheaper successful request is what matters; an inexpensive answer that forces repeated corrections can become the expensive option.

05 / Keep repeated document questions inside the budget

CheapRelay provides pay-as-you-go access to both models through one API key, with no monthly fee or subscription. Billing follows real token usage, including cache reads and writes. Compare the current model rates there and sign up for CheapRelay when you are ready to run your small test.

For repeated questions, check whether caching is supported and how its charges apply. Do not assume cached input is free. Record billed usage after each request, avoid resending unnecessary history and stop when the next estimated request would exceed your remaining budget. Your own small receipt is a better purchasing guide than an unsupported claim about which model handles long documents best.