Can ChatGPT read PDFs?
Yes, on every plan including Free. PDF is a supported upload type and ChatGPT will summarise, search and extract from it. The limit that catches people is not size: outside ChatGPT Enterprise, uploads are text-based retrieval only — ChatGPT extracts the digital text and discards any images. A scanned document has no digital text, so it arrives effectively blank.
The answer is yes. The useful answer is what happens to a scan.
PDF sits in OpenAI's supported list alongside XLSX, XLS, CSV, TSV, DOCX, PPTX and TXT, and file uploads are available on Free and paid plans alike. The documented uses map exactly onto what people want from a PDF: summarising a long paper in plain terms, finding every reference to a topic, pulling quotes, extracting metadata such as author and creation date, and comparing two documents against each other.
Then there is the sentence that decides whether your particular PDF works. Asked whether it handles images embedded in documents and PDFs, OpenAI answers that ChatGPT Enterprise supports Visual Retrieval for PDF files, and that "All other plans and document files only support text-based retrieval. This means that ChatGPT will extract digital text from the file and discard any images."
Read that against a real scanned document — a contract someone photographed, an old paper, a receipt — and the consequence is stark. A scan is a picture of text, not text. Outside Enterprise, that picture is discarded, and the model is left summarising a file it cannot see. The failure is quiet: you get an answer, and the answer is not grounded in the document.
The size limits are generous by comparison and rarely the problem: 512 MB hard ceiling per file, 2M tokens for text and document files, up to 80 files every 3 hours, and 3 uploads per day on Free. Scope: we did not run any of this — no ChatGPT account here — so every claim above is OpenAI's own, read on the date shown.
Making a difficult PDF work.
Thirty-second test, no tools required: open the PDF and try to select a sentence with your cursor. If the text highlights, it is digital text and ChatGPT will read it. If your cursor draws a box over the page instead, it is a scan, and outside Enterprise the content will be discarded on upload.
This is worth doing before you ask a question about the document rather than after, because the failure does not announce itself — you get a fluent answer built on whatever context surrounds an empty file.
OCR it first. Run the file through any optical-character-recognition tool — most operating systems, PDF readers and cloud drives now do this — and upload the searchable version. This is the reliable fix and it costs nothing.
Or upload the pages as images. Image inputs work on every plan and all ChatGPT models accept them, so screenshotting a handful of pages turns a scan into something the model genuinely sees. It does not scale past a few pages, and OpenAI's own limitations list warns that the model does worse on non-Latin scripts, rotated text and images where meaning depends on line style or colour.
The 2M-token ceiling per document is high enough that most books clear it. The practical constraint is different: a summary of a 400-page PDF is a summary, and detail gets lost in ways that are invisible from the answer.
What works better is extraction over summarisation — asking for every mention of a specific term, pulling the quotes that bear on one question, or comparing two documents on a named dimension. Those are the tasks OpenAI's own documentation lists first, and they are the ones where you can check the output against the source.
If you want to know how large a document is in tokens before you upload it, our token counter will tell you.
Uploaded files are retained for as long as the chat that holds them, and are removed within 30 days of deleting that chat, your account, or the custom GPT they were attached to — with the usual carve-outs for de-identified data and legal retention. Files also count against your Library storage, which is 500 MB on Free.
If the document is confidential, the relevant control is Temporary Chat: files uploaded there are never saved to Library, and the conversation is not used for training. Our page on chat privacy covers what the defaults are before you change them.
For scans, tables and hundreds of pages.
ChatGPT handles ordinary digital PDFs well. These are the routes when it does not. We have not benchmarked extraction accuracy across any of them.
Frequently asked.
Quick follow-ups people search after this question.