Claude Can See Your PDFs Now and It's Weirder Than You Think

Claude can now process PDFs visually — not extracted text, not OCR'd strings dumped into a context window, but the actual rendered pages, treated like images.

The limit is 32MB and 100 pages. Which sounds restrictive until you run a 200MB PDF through Acrobat's compress tool and watch it come out at 28MB. Turns out most of the bloat in a large PDF is cruft Acrobat will happily discard without touching the content you care about.

The thing that took me a minute to understand is that you don't point Claude at 77 PDFs. You RAG across 77 PDFs to find the ones that matter, then you hand Claude the relevant few with visual processing turned on. The retrieval step is still retrieval — you're not asking Claude to hold an entire document corpus in its head. You're using it as the sharp end of the process, not the index.

Once you've got the right document in front of it, the experience is seamless. The user doesn't see the seam between "found it" and "read it." It just works.

I turned it on in the preview settings and ran some RVA PDFs through it — the kind of dense municipal documents that eat text extraction alive, where the layout is doing half the work and a naive parse turns a table into word soup. Claude handles them the way a person would handle them. It looks at the thing.

That's the part that's bonkers. Not the compression trick, not the RAG pipeline. The part where the model is actually looking at your document the way you look at your document, and it turns out that matters enormously for anything where structure and visual hierarchy carry meaning.

Which is most documents.