$0.036 for 200 Textbook Flashcards: A Sample Receipt

01 / The 3.6-cent receipt starts with assumed rates

Turning textbook passages into flashcards can cost pennies in model usage. But a useful estimate needs more than a cheap headline: it needs input tokens, output tokens, review passes, and a clear boundary around what is excluded.

Here is a worked budget for 200 finished flashcards. Every token quantity below is a planning assumption, not a measured benchmark. The rates are illustrative: $0.20 per million input tokens and $0.80 per million output tokens. They are not a quote for a particular model or a claim about current CheapRelay pricing. Check your selected model’s rates before running the job.

• ■ First-pass input: 80,000 tokens × $0.20 / 1,000,000 = $0.0160.
• ■ First-pass output: 16,000 tokens × $0.80 / 1,000,000 = $0.0128.
• ■ Review-pass input: 20,000 tokens × $0.20 / 1,000,000 = $0.0040.
• ■ Review-pass output: 4,000 tokens × $0.80 / 1,000,000 = $0.0032.
• ■ Total model usage: 120,000 tokens, costing $0.0360.
• ■ Budget per finished card: $0.00018, assuming 200 survive review.

That is 3.6 cents in estimated API usage, not the total cost of studying. Text extraction, paid OCR, software, taxes, and your editing time are outside this receipt.

02 / Build the input budget around excerpts, not page counts

Pages are a poor billing unit. A page of equations, a dense history chapter, and a scanned diagram produce very different extraction results. Work from selectable text, preserve section headings and page references, and measure tokens before submitting requests.

For this example, allocate 60,000 input tokens to textbook excerpts and 20,000 to instructions, formatting requirements, and other request overhead across ten batches. That produces the 80,000-token first-pass allowance. Each batch targets roughly 20 cards, with an average output allowance of 80 tokens per card, including field labels and source references.

These are budget targets, not guarantees. Long explanations, duplicated passages, or verbose JSON can push usage higher. Avoid sending the entire textbook with every batch. Send the relevant section and enough surrounding context to resolve definitions and exceptions.

03 / Ask for cards that can be checked against the textbook

A cheap generation pass is wasted if the cards teach unsupported claims. Ask for one testable fact per card, a short answer, and a source page or section. Require the model to skip material it cannot resolve from the supplied excerpt.

A practical instruction is: Create up to 20 flashcards from this excerpt. Return question, answer, source reference, and a short supporting quotation for each. Test one concept per card. Preserve qualifications and units. Do not add outside facts. Skip ambiguous claims.

Keep supporting quotations brief and use material you are authorized to process. Remove those quotations from the study-facing export if they would reveal answers or clutter review. For equations and tables, inspect the extracted text manually before generation; missing symbols can produce confidently wrong cards.

04 / Reserve $0.0072 for a bounded review pass

The review allowance covers 20,000 input tokens and 4,000 output tokens. Send draft cards with their relevant source snippets, rather than resending every chapter. Ask for flagged errors, duplicate concepts, vague prompts, and corrections only. A full rewrite would consume a different output budget.

Human-check the flagged cards and spot-check unflagged ones. Reject unsupported answers instead of paying for repeated guesses. If only 160 cards survive, the same $0.036 receipt becomes $0.000225 per usable card. Quality changes the denominator.

05 / Keep the actual bill beside the study export

CheapRelay offers pay-as-you-go API access with no monthly fee or subscription. Sign up for CheapRelay when you are ready to test one small textbook batch, then compare its recorded usage with this budget before scaling.

CheapRelay bills real tokens, including cache reads and writes. This example assumes no cache activity; if your requests use caching, account for those categories at their applicable rates. Save request usage, rejected-card counts, and the final export together. Your useful receipt is actual spend divided by cards you trust—not simply cards the model generated.