2026-09-17

Why won't ChatGPT, Claude, or Gemini take your whole PDF? The real upload limits — and how to trim or combine it first

Handing a whole PDF to an AI assistant — "summarize this lease," "what does page 40 of this research paper actually claim," "pull the numbers out of this scanned tax return" — has become one of the ordinary ways people get through a long document now. What doesn't get talked about as much is that ChatGPT, Claude, and Gemini each enforce a real, different, and surprisingly easy-to-miss ceiling on what one upload can contain, and hitting it doesn't announce itself as "this file is too big." It shows up as an upload that's flatly rejected, one that goes through but the assistant's answer is confidently wrong about a part of the document it never actually saw, or one that just won't attach at all. That's a mechanical problem with a mechanical fix, and none of it requires paying for a bigger plan.

The actual ceilings — not vibes

Each assistant publishes its own numbers, and they don't match each other:

• Claude. Anthropic's own PDF-support documentation states it plainly: a maximum request size of 32 MB, and a maximum of 600 pages per request — except that ceiling drops to 100 pages the moment the conversation's context window is under 1M tokens, which is the default for most ordinary use. The same requirements table has a second rule that's easy to miss entirely: the file has to be a "Standard PDF (no passwords/encryption)" — an encrypted PDF isn't a large-file problem at all, it's a flat rejection no matter how small it is.
• ChatGPT. OpenAI's own file-uploads FAQ sets the size cap generously, at 512 MB per file — but the number that actually bites first for a text-heavy document is a separate one: content is capped at 2,000,000 tokens per file, and a long report can sail past that while still weighing only a few megabytes. The same FAQ caps how often you can upload at all — a free account gets 3 file uploads a day.
• Gemini. Google's own Gemini Apps help page caps every uploaded file, PDF included, at 100 MB, and limits a single prompt to at most 10 files.

None of these are secondhand guesses — they're each vendor's own current published limit, and each vendor is free to change theirs without much notice, so treat "current as of September 2026" as exactly that, not a permanent number.

Two failures that feel the same but aren't

The size/page ceiling and the encryption rule produce two different-looking failures that both read as "it won't take my PDF." A file that's too big or has too many pages either gets rejected outright or — worse — gets accepted and then silently truncated, so the assistant answers about the part it actually received with no way to tell you it never saw the rest. A password-protected PDF is a different failure entirely: Claude's own documentation excludes it categorically, regardless of size, so a two-page encrypted statement fails for the same reason a six-hundred-page one would. Knowing which failure you're looking at decides which fix applies — you can't split your way out of an encryption rejection, and you can't unlock your way out of a page-count rejection.

Fix #1: trim it down to the part the question is actually about

If the real ask is "what does the arbitration clause in this 80-page lease say," an assistant doesn't need to see all 80 pages — it needs to see the handful that matter, comfortably inside any platform's ceiling. NearPDF's split tool doesn't have a page-range picker (it produces one PDF per page, zipped, and nothing else), but its own documented workaround for exactly this is the delete tool: add the file, click the ✕ on every page outside the range you actually need, and save — the result is one trimmed PDF containing just that range. That's the identical technique already covered in more depth in NearPDF's email-size-limit post, for a different reason there (an inbox cap instead of an AI upload cap) — the mechanism doesn't change, only what you're trimming it for.

Trimming buys you more than headroom under a ceiling. A 2023 paper from a Stanford-led research team, published in the Association for Computational Linguistics' TACL journal as "Lost in the Middle: How Language Models Use Long Contexts," found that these models are measurably better at using information placed near the start or end of what they're given than information buried in the middle of a long context. That's a general, peer-reviewed finding about how the underlying models behave — not a claim about any one assistant's current behavior, and not a NearPDF guarantee about any specific answer — but it's a real reason a shorter, on-topic file can get you a better answer, not just a smaller one.

Fix #2: one ordered file beats five separate uploads

The opposite problem shows up just as often: three years of the same form, a syllabus plus its grading rubric, a contract plus its signed addendum — several small, related PDFs you want one assistant to consider together, rather than as separate uploads it may read in whatever order it received them. NearPDF's merge tool handles the plain case: add the files in the order you want, save one combined PDF. Its organize tool has a detail worth knowing for anything more particular: add several PDFs to it at once and every page from every file lands in one shared grid, so dragging pages into a specific order — interleaving one file's pages with another's, not just appending file B after file A — quietly doubles as a merge with full control over the result.

The password wall isn't an AI-specific rule

Claude's own requirements table excludes encrypted PDFs outright, but that's not a restriction any AI vendor invented for this occasion — it's the identical wall NearPDF's own tools already run into. NearPDF's merge tool checks every file you add and skips a password-protected one automatically: "It's skipped automatically — NearPDF can't open encrypted PDFs, so remove the password first," in the tool's own words. If a bank portal or scanning app handed you a locked PDF, that password has to come off in whatever created it, or in a PDF reader that supports removing one — NearPDF has no decryption feature, and neither trimming nor merging changes that.

The honest limits

NearPDF doesn't read, summarize, or understand anything inside your PDF, and it never uploads a file anywhere — to an AI service or otherwise — every trim, merge, reorder, and rotation described here happens entirely on your device, before you decide whether to send anything to anyone. It also doesn't know what any assistant's limit will be next month; the numbers above are each vendor's own current documentation, not a promise NearPDF is making on their behalf. What it can do is the preparation step in between: turning one oversized or over-the-page-limit PDF into the specific range that actually matters, or turning several related PDFs into the one ordered file that gives an assistant the clearest shot at a useful answer — without a single page of it leaving your browser until you choose to send it somewhere yourself.