Microsoft's free MarkItDown turns your files into plain text. Anthropic's own guide to building skills for Claude costs 64,141 tokens as a PDF. As text: 12,987. Almost five times fewer.
Made by Microsoft: MarkItDown, free and open source. I measured it on five real PDFs and tested it in Claude Code, start to finish.
On your phone? MarkItDown runs on a computer. Open federicodonatone.com/pdf there, or .
📄
skills-guide.pdf33 pages, by Anthropic
64,141tokens
↓ MarkItDown, free
📝
skills-guide.mdthe words, no pictures
12,987tokens
4.9× fewerSame modelAnthropic's token counter
Proof · 28 Oct 2026
One real PDF, measured
Anthropic's own guide to building skills for Claude, 33 pages. I asked Anthropic's token counter what it costs Claude Opus 5.5: once as the PDF, once as MarkItDown's text.
64,141tokens as a PDF
12,987tokens as text
4.9×fewer tokens
Page by page
Each of the 33 pages, counted on its own.
As a PDFAs text
Every page: 1,583+
Page 1Page 33
As a PDF, no page costs less than 1,583 tokens, even the back cover with one word on it. As text, that page costs 6 tokens. Counted one by one, the pages add up to 64,973 as PDFs, a little more than the whole file, because each page went as its own small PDF.
Five real PDFs
Same counter, same model, same day.
The Complete Guide to Building Skills for Claude33 pages · Anthropic
4.9×
PDF64,141
Text12,987
A practical guide to building agents34 pages · OpenAI
6.2×
PDF62,990
Text10,080
Berkshire Hathaway's 2025 letter to shareholders20 pages · Berkshire Hathaway
2.9×
PDF48,496
Text16,508
Bitcoin: A Peer-to-Peer Electronic Cash System9 pages · Satoshi Nakamoto
2.8×
PDF21,287
Text7,555
Attention Is All You Need15 pages · Ashish Vaswani and co-authors
2.0×
PDF38,354
Text19,460
My video says five times fewer. That's this guide: 4.9×. The other four real PDFs gave 2.0× to 6.2×. Your PDF will have its own number.
How I measured. Anthropic's count_tokens API, model Claude Opus 5.5, 28 Oct 2026. The PDF went as a document, the way Claude gets it. The text went as plain text: MarkItDown 0.1.9's output, byte for byte the file my guide below makes. The short question sent with both (15 tokens) is taken out of both numbers.
Why a PDF costs so much
Claude takes a picture of every page
When you send a PDF, Claude gets the text of each page and an image of that page too. The image costs tokens even when the page is almost empty. MarkItDown keeps the words and drops the pictures.
Anthropic's PDF docs, word for wordThe system converts each page of the document into an image.
Anthropic's Help Center, on the Claude appsClaude analyzes both text and visual elements (like images, charts, and graphics) in PDFs of 100 pages or fewer.
Back cover · page 33“claude.ai”1,583 as a PDF6 as text
Chapter page · page 7Chapter 2: Planning and design1,592 as a PDF15 as text
Full page of text · page 9The busiest page of the guide2,367 as a PDF794 as text
Every number above comes from the same page-by-page count. The words on a page are cheap. The picture of the page is the expensive part.
Not only PDFs
It also reads Word, Excel, PowerPoint and voice notes
I made four small test files and converted each one in a clean sandbox on a Mac. This is exactly what MarkItDown wrote. Headings, lists and tables stay, so Claude still sees the structure.
📝
Word · plan.docxA short plan
MarkItDown wrote:# Q4 Launch Plan
Goal: ship the new pricing page by Friday.
* Owner: Maria
* Budget: $12,000
* Launch: Nov 14
MarkItDown wrote:### Audio Transcript:
meeting notes we agreed to ship the new pricing page on Friday and Maria will send the proposal to the client on Monday
Also on its list:ImagesHTMLCSV, JSON, XMLZIP filesEPubI tested the six in my video.
The best one
Paste a YouTube link. Get the whole video as text.
MarkItDown reads the video page and its captions. I tried the very first video on YouTube, “Me at the zoo”, 19 seconds long.
What MarkItDown wrote, first lines:# YouTube
## Me at the zoo
### Video Metadata
- **Keywords:** me at the zoo,jawed karim,first youtube video
- **Runtime:** PT0M19S
(its description)
### Transcript
All right, so here we are, in front of the…
Tested on 28 Oct in a clean sandbox at 15:26 UTC, and in Claude Code at 14:01 UTC: Claude ran MarkItDown on the link and answered from the transcript.
It needs captions. No captions, no transcript. And YouTube blocks these requests at times: on the day of my test it blocked mine for a while, and MarkItDown returned only YouTube's footer links. Waiting fixed it.
Read this first
When not to convert
Scanned PDFs. A scan is only pictures, so MarkItDown finds no words. I made a scanned copy of one page: MarkItDown returned nothing. Drag scans into Claude as they are.
Charts and diagrams that matter. MarkItDown keeps the words and drops the pictures. In the famous Transformer paper, the caption of Figure 1 stays. The diagram is gone.
Long PDFs in the Claude apps. Anthropic says that above 100 pages the apps read the text only and skip the pictures.
The honest fine print
Five times is one PDF. My video's 4.9× is Anthropic's guide. The other four real PDFs gave 2.0× to 6.2×.
I measured tokens. Fewer tokens leave Claude more room. I did not measure whether the answers get better.
Messy layouts. Columns and side boxes can get mixed, and some PDFs lose their spaces: the Transformer paper gave TheTransformer-modelarchitecture. In my test, Claude's summary of the converted guide still covered the same points.
Voice notes go to Google. MarkItDown sends the audio to Google's free speech service and listens for English. mp3 and m4a files also need ffmpeg. WhatsApp voice notes (.opus) are not on its list.
Your files stay yours. PDF, Word, Excel and PowerPoint are converted on your computer. Voice notes and YouTube links use Google's and YouTube's servers.
One Mac, one day. I tested MarkItDown 0.1.9 on a Mac on 28 Oct 2026. I did not test Windows.
✦ Free · link and setup guide
Get the link and my setup guide
📄📽️📝📊🖼️🎙️🌐🧾🗂️▶️📚
The link to MarkItDown, with a copy button. Free, MIT, by Microsoft.
The install, in one line. I tested it today in four clean sandboxes on a Mac: about 20 seconds.
One command for each file: PDF, Word, Excel, PowerPoint, voice notes, YouTube, and a whole folder at once.
My Claude Code setup: one line and one permission, and Claude converts your files before it reads them. Plus what to do on claude.ai and in Claude Desktop.
You're also joining AI Frontier, my weekly email: one email every Thursday, unsubscribe anytime. Privacy
Step 1 · 2 minutes
Install MarkItDown in one line
It runs on your computer. Free, no account, no API key.
Locked. The link, the install line I tested on a clean Mac, and the fix if your terminal says it can't find it.