If your workflow involves pasting PDFs, Word docs, spreadsheets, or screenshots directly into a chat with an LLM — asking it to "read this" or "summarize this" — you're paying input-token prices for work a deterministic parser could do for free. Here's the actual math, and how to avoid it.
When you drop a PDF or image into Claude or ChatGPT, the model has to spend tokens on two things: understanding the raw structure (layout, tables, OCR for scanned pages or images) and whatever you actually asked it to do with the content. The extraction step is pure overhead — it doesn't require reasoning, just correct parsing — and every model provider charges the same input-token rate for it as for the reasoning you actually wanted.
SlimdownPDF runs a tiered pipeline before ever touching an LLM:
| Step | Cost | Handles |
|---|---|---|
| Native parsers (pandas, openpyxl, python-docx, html2text) | $0 — 0 tokens | CSV, Excel, Word, HTML, JSON, XML |
| Local Tesseract OCR | $0 — 0 tokens | Images, most scanned PDFs |
| LLM fallback (optional) | Real token cost, only when needed | Low-confidence OCR, unusual layouts |
Most everyday documents — spreadsheets, Word reports, web pages, JSON exports — never reach the LLM step at all. The ones that do (scanned handwriting, complex mixed-layout PDFs) are the actual edge cases, not the default.
Take a 10-page report that converts to roughly 8,000 characters of clean Markdown (~2,000 tokens). Pasted directly as input to a few current models (official pricing, verified July 2026):
| Model | Input price / 1M tokens | Cost for ~2,000 tokens |
|---|---|---|
| Claude Opus 5 | $5.00 | $0.010 |
| Claude Sonnet 5 | $2.00 | $0.004 |
| GPT-4o | $2.50 | $0.005 |
| GPT-4o mini | $0.15 | $0.0003 |
A single document looks cheap. The cost adds up when it's every document, every day, across a team — and it compounds further for scanned/image-heavy files, which use vision tokenization and cost meaningfully more than plain text per page.
Convert first, with a tool that tries the free path before the paid one, then feed the clean result — not the raw file — into whatever LLM workflow you actually need tokens for. That's the entire idea behind SlimdownPDF: PDF, Word, Excel, PowerPoint, images, audio, HTML, CSV, JSON, XML, ZIP archives and YouTube URLs in, clean Markdown out, and a live estimate of what each result would have cost across Claude and GPT models so you can see the difference yourself.
Convert a file free →