5 ways to cut AI API costs when processing PDFs

If you are running PDFs through an LLM API at any volume, the bill adds up faster than you would think. The good news is that most of the cost is avoidable. A lot of what you send to the model is noise the model does not need.

Here are the levers that actually move the number, roughly in order of impact.

1. Pick the right model for the task

This is the biggest lever by far. Flagship models cost 10 to 20 times more than their mini or haiku variants. For extraction, classification, and simple summarization, the cheaper models are usually good enough.

Reserve the expensive models for tasks that genuinely need deeper reasoning. Run a small test batch on both and compare the output quality before assuming you need the flagship.

2. Strip the noise before you send

Raw PDF extraction carries a lot of dead weight: repeated headers and footers on every page, page numbers, boilerplate legal text, navigation artifacts. None of it helps the model, and all of it costs tokens.

Converting to clean Markdown and trimming repeated elements can cut 15 to 30% off the token count on a typical business document. On a large pipeline, that is a direct percentage off the bill.

3. Send only the pages you need

If your question is about the payment terms in a contract, you do not need to send the other 40 pages. Extract the relevant section first and send just that.

This sounds obvious, but a surprising number of pipelines send the whole document every time out of convenience. For repeated queries against the same long document, splitting once and reusing the relevant chunk saves on every call.

4. Chunk smart, not just small

If you are building retrieval, chunk size affects cost. Retrieval sends the matched chunks to the model, so oversized chunks mean you pay for context the query did not need. Undersized chunks mean you retrieve more of them to get the same coverage.

Tune chunk size to your actual retrieval pattern. For most document QA, 512 to 1024 tokens per chunk is a reasonable balance between precision and cost.

5. Cache what repeats

Some providers offer prompt caching for repeated context. If you send the same system prompt or the same document header on every call, caching can cut the cost of that repeated portion substantially. Check whether your provider supports it and structure your prompts so the stable part comes first.

Measure, then optimize

You cannot optimize what you have not measured. Before restructuring a pipeline, run a representative document through the Token Cost Calculator to see the baseline cost across models. Then apply the levers and measure again.

For the cleanup and splitting steps, the PDF to Markdown converter trims noise, and the PDF Chunker handles the retrieval segments. Both run locally, so you can optimize sensitive documents without sending them anywhere first.