Why PDFs and spreadsheets eat your AI limits

When an AI reads a PDF itself, each page often goes in twice. Spreadsheets lose their formulas. Here's what's happening, and how to get the same answers for far fewer tokens.

REPORT.PDF TOKENS A PAGE AS PAGES CLEAN TEXT

You upload a ten-page PDF, ask one question, and a chunk of your daily limit is gone. Or you share a spreadsheet and the AI confidently explains numbers that were really formulas. Neither is bad luck. It's how most AI tools read files.

PDFs: every page, twice

When you upload a PDF to Claude or ChatGPT, it typically reads each page two ways: the text, plus a picture of the page so it can see charts and layout. The picture is the expensive part. Anthropic's own documentation gives a sense of scale: a 3-page PDF read as pages costs about 7,000 tokens, while the text alone is about 1,000.

Reading the PDF as pages

PDF page→Picture of the page+Its text

Every page goes in twice.

Reading a clean version

PDF page→Text + tables, in order

Once, with the structure kept.

In our own tests on sample files (a text report, a booking confirmation, an annual report page with a chart, a board pack with a figure), reading the PDF directly came to roughly 2,200 tokens a page. A clean text version with the tables rebuilt came to roughly 650.

Fewer tokens with a clean version, sample files

5-page text report71%Mostly text: 71% fewer tokens
2-page booking confirmation71%Short, text-heavy: 71% fewer
Annual report page with a chart80%A chart makes the page picture costlier: 80% fewer
2-page board pack with a figure92%Mostly visuals and white space: 92% fewer
SampleFewer tokens
5-page text report71%
2-page booking confirmation71%
Annual report page with a chart80%
2-page board pack with a figure92%
Hover or tap a bar for detail.

Spreadsheets: the AI sees numbers, not the model

A spreadsheet is more than its values. It's the formulas, the inputs they point to, named ranges and how one sheet feeds the next. When an AI opens an Excel file itself, it usually writes and runs code to read it first, often more than once, and that code mostly pulls out values. The logic disappears, so the AI can't tell you why a number is what it is. It guesses.

Who this matters to

  • Students and anyone on a free plan. Free plans meter how much the AI reads, or cap uploads a day. A lighter file means more questions before you hit the limit.
  • Teams and firms. When a whole team reads reports, decks and models through AI every day, the difference per file becomes a real line in the budget. Many companies now set per-person limits on AI use, so every token saved is room for actual work.
  • Anyone sharing sensitive files. The less you upload, the less you expose. Removing emails, phone numbers and account numbers before a file leaves your computer is a good habit regardless of tool.

What to do about it

  1. Send text, not pages, when the content is mostly text. Many tools let you paste text or upload a text file instead.
  2. Rebuild tables properly as Markdown tables, so columns stay in order.
  3. Give the AI the formulas, not just the values, when you want help with a model.
  4. Strip personal details before uploading.

The shortcut

We built Unpack to do all four in one step. Right-click a file on your Mac, or open it on Windows (beta), and you get one clean Markdown file any AI reads straight away: text and tables in order, every spreadsheet formula and how the sheets connect, and private details swapped for labels like [EMAIL 1]. It runs on your own computer, so your files never leave it. Try it free for 7 days, then it's a one-time purchase.