Is it cheaper to send a PDF to an LLM or extract the text first? (RAG question)

Extracting the text first is usually cheaper, because you send only text tokens. Sending the PDF directly usually costs more, since models like Claude, ChatGPT and Gemini also process each page as an image, but it keeps screenshots, charts and diagrams that text extraction loses.
A few days ago, while working on a RAG project, one simple question stopped me:
“What costs more — sending the PDF directly to an LLM, or extracting the text first?”
Sounds simple… but then came the next question: What if the PDF has screenshots, charts, or diagrams?
That took me down an interesting rabbit hole of how Claude, ChatGPT & Gemini process PDFs — and why cheaper input isn’t always the better choice.
Sharing what I learned in this visual
What would you choose for RAG — PDF or extracted text?
Ebooks to watch for--https://www.aiplaybookhub.com/books/genai-testing-for-qa.html


