$2.78 for 10,000 Documents: A Lean API Budget

01 / The $2.78 estimate starts with a narrow job

Processing 10,000 documents does not automatically require a large AI budget. For a small extraction task, the estimated model bill can be $2.78. That is not a quote for arbitrary PDFs: it assumes 2,000 total input tokens and 100 output tokens per document, using the listed Qwen3.7 Plus rates explained below. Long reports, scans, retries, and extra reasoning tokens change the bill. ([cheapapirelay.com](https://cheapapirelay.com/))

Start by defining one useful result. For example: extract a supplier name, invoice date, currency, and total. Do not request a full summary alongside those fields unless someone actually needs it. The cheapest useful workflow is the one that sends only necessary material and produces an answer you can validate.

02 / Extract text before paying for interpretation

Create three folders: originals, extracted text, and results. Keep originals unchanged. Assign each file a stable document ID, and maintain a CSV manifest with its filename, processing status, and result location.

For text-based PDFs, try local text extraction first. PyMuPDF supports plain-text extraction, although the resulting reading order depends on how the document was created. Inspect a few examples before trusting the output, especially invoices with columns. ([github.com](https://github.com/pymupdf/PyMuPDF/blob/main/docs/app1.rst?utm_source=openai))

Use optical character recognition, or OCR, for scanned pages that need it. Save that extracted text rather than repeating OCR on every attempt. PyMuPDF's documentation specifically recommends reusing OCR results because OCR is substantially slower than ordinary text extraction. Local processing still consumes your computer's time and electricity. ([pymupdf.readthedocs.io](https://pymupdf.readthedocs.io/en/latest/recipes-ocr.html?utm_source=openai))

• ■ Skip exact duplicate files after checking their contents or hashes.
• ■ Extract straightforward fields locally when reliable rules are sufficient.
• ■ Keep relevant pages and surrounding context; do not blindly truncate.
• ■ Review whether you have permission to send the documents to an external service.

03 / Build the receipt from tokens, not file counts

At the time of checking, CheapRelay lists Qwen3.7 Plus at $0.1208 per million input tokens and $0.3625 per million output tokens. Treat these as a checked price snapshot, not a permanent promise. ([cheapapirelay.com](https://cheapapirelay.com/))

Here is the arithmetic for our hypothetical workload. The 2,000 input tokens include both instructions and document text; the 100 output tokens represent the entire billable output assumption:

• ■ 10,000 × 2,000 input tokens = 20 million input tokens.
• ■ 20 × $0.1208 = $2.416 for input.
• ■ 10,000 × 100 output tokens = 1 million output tokens.
• ■ 1 × $0.3625 = $0.3625 for output.
• ■ Total estimated model usage: $2.7785, rounded to $2.78.

This calculation excludes separate OCR services, storage, taxes, retries, and any additional billable token categories. It assumes no cache activity. CheapRelay prices input, output, cache writes, and cache reads separately, so reconcile your estimate against recorded usage rather than treating short visible answers as proof of low cost. ([cheapapirelay.com](https://cheapapirelay.com/))

04 / Spend pennies on a 50-document pilot first

Choose 50 representative documents: short, long, clean, scanned, and awkward. Manually record the correct fields. Ask the model for those fields only, with null for missing values and no explanatory essay. Tell it to treat document text as data, not instructions.

Check every pilot result against the originals. Measure field accuracy and actual billed tokens. Multiply average cost per document by 10,000, then add an explicit contingency. If important totals or dates are wrong, fix extraction or instructions before scaling.

05 / Run small batches that can resume after failure

Process one document per request initially. Save each successful result immediately, together with its document ID, token usage, and cost. Mark failures separately. Restart from unfinished IDs instead of rerunning the whole folder.

Set a retry limit of two and pause when your own spending ledger reaches its budget threshold. Start with low concurrency. Validate required fields, date formats, and numeric totals before accepting results; send questionable records to a review queue.

06 / Buy only the model usage your pilot justifies

CheapRelay offers pay-as-you-go token billing without a monthly fee, making it suitable for a one-off experiment without a recurring commitment. Sign up for CheapRelay when you are ready to run your measured pilot. ([cheapapirelay.com](https://cheapapirelay.com/))

The practical target is not a magical price per file. It is a repeatable workflow: extract locally, test quality, measure tokens, save progress, and pay again only when another attempt is justified.