PDF to Markdown for ChatGPT & Claude: the measured token savings

7 August 2026 · Milldown team · All posts

Every converter site repeats the same claim: “Markdown uses 30–60% fewer tokens.” Few say where the number comes from, and none break it down by format. We did the measurement — with the o200k tokenizer, on real documents — and the truth is more interesting: some formats save far more than 60%, and one format saves nothing at all.

The numbers, per format

Source formatCompared againstMeasured token saving
PDFNative PDF upload (text + per-page image)60–75%
Web page / HTMLRaw HTML source53–66%
Word (.docx)The file’s underlying XML97%+
PowerPoint (.pptx)The file’s underlying XML97–99%
Excel (.xlsx)CSV of the same sheetNegative — see below

Methodology: token counts use the o200k tokenizer (the family used by current frontier models), computed locally on a corpus of real reports, contracts, decks, and articles. Milldown shows the same measurement live for every file you convert, so you never have to take a blog post’s word for it — including this one.

Why PDF is the headline number

Uploading a PDF natively bills every page as extracted text plus a page image — roughly 1,500–3,000 tokens per page in Claude’s documented pricing. Once the text is out, that per-page image cost is pure overhead, and it’s re-billed on every conversation turn. Sparse pages save the most, because the fixed image cost dominates. Details in our PDF to Markdown guide.

Why Office formats look absurdly high

A .docx or .pptx file is a zip of XML where a single bold word can cost hundreds of characters of markup. Against that raw XML, Markdown’s saving is 97%+ — real, but only the fair comparison if your pipeline would otherwise ingest the raw file contents (as many RAG loaders do). Chat apps that extract text server-side before the model sees it capture some of this saving for you; what they don’t do is preserve heading structure and tables as cleanly, or let you edit the result before sending it.

The honest exception: spreadsheets

Markdown tables are more expensive than CSV — all those aligned pipes and dashes add up. On our test sheets, CSV came out about 32% smaller than the equivalent Markdown table. That’s exactly why Milldown has a dedicated CSV output mode for Excel files: the right target format depends on the source. Use Markdown tables when a human will read them; use CSV when a model will.

One more honesty note: if you call the API with prompt caching, repeat reads of the same document get cheaper, so dollar savings shrink on later turns (roughly 3–5× smaller). Context-window savings don’t shrink — a smaller document leaves more room for your actual conversation, cached or not.

Try it on your own documents

Claims are cheap; measurements are better. Drop your own files into Milldown and the token badge shows the before/after for each one — computed on your Mac, with nothing uploaded.

Measure it yourself

Convert PDFs, Office files, and web pages to clean Markdown and see the exact token saving per file — entirely on your Mac. Free for 14 days.

Get Milldown for macOS

macOS 14+ · Apple Silicon · $39 one-time after trial (excl. VAT) · no subscription