You've done this before: drag a report into ChatGPT or Claude and ask it to summarize, rewrite, or answer questions about it. What you don't see is what happens in the half-second between "upload" and "reading."
That half-second is where the token cost gets decided — and it depends entirely on what kind of file you just handed over.
Uploading a .md file: nothing happens, because nothing needs to
A Markdown file is plain text. When you upload it, the AI reads it exactly as it is — the same characters you'd see in a text editor. There's no format to unpack first. What gets tokenized is your actual content, and nothing else.
Uploading a .docx file: it has to be figured out first
A .docx isn't text — it's a zipped bundle of XML files describing styles, formatting, and structure, wrapped around your content. Before an AI can read a single sentence, something has to open that archive, walk through the XML, and pull the actual words back out.
That extraction step is expensive in two different ways:
Enjoying the article?
Create a free account to finish reading and join the conversation.