How to upload images and PDFs, and which models accept what

When the paperclip refuses a file, the file is usually fine. Check whether the selected model accepts that file type.

MultiChats carries 59 models from 17 providers and they do not accept the same attachments. The registry check on 19 October 2026 found: 46 take images, 24 take PDFs, 6 take video. Those 24 PDF models all sit inside the 46, so no model in the catalogue reads a PDF but refuses a photograph. Thirteen models take no image of any kind.

Your plan matters too, and it is a separate check. Some file types are gated by your subscription rather than by the model, and both gates have to open before a file will attach.

Where the app says what a model takes

Three places, in increasing detail.

The model picker carries a row of filter chips above the list, and two of them narrow it to models that will take a given file: Image Upload and PDF Upload on the web, Vision and PDF analysis in the mobile app. Neither platform offers a video chip, so video is the one capability you cannot filter for.

Every model in that list carries capability markers, green for images, blue for PDFs, rose for video, and each is drawn only when the model has that capability, so a missing marker is the answer. Hover one on the web and the tooltip spells it out: "Supports image upload and analysis", "Supports PDF document upload and analysis", "Supports video upload and analysis".

On the web, hovering the small info button on a card brings up the fuller model info card, which lays the same capabilities out as a labelled grid next to the model's context window and any privacy notice it carries.

One thing the composer deliberately does not do is filter the file dialog. The dropzone accepts every type you can drop on it, so that a refusal can be explained instead of leaving a file to vanish without a word. Every rejection below arrives as a message on screen. The wording quoted in each entry is the browser's; the app says the same thing in a shorter form.

Images, on 46 of the 59 models

• Models: 46 of 59, including 14 of the 22 available without a subscription.

• Plan: any, free included.

• What reaches the model: the image file itself.

• If the model cannot take it: "This model can't read images. Switch to a model with vision to send images."

• If your plan cannot: not a plan limit. Every tier can attach an image.

• Verdict: the safest thing to attach, and the widest capability we have.

iPhone photos are the wrinkle. iOS captures HEIC by default, and most desktop browsers have no HEIC decoder, so a photo that looks fine in your camera roll can turn up as a broken thumbnail when you open the same conversation on a laptop. Since early September the mobile picker re-encodes HEIC and HEIF to JPEG at pick time, before the upload starts, and corrects the recorded file size. That fix lives in the app code rather than in every installed build, so if your photos still go up as HEIC, your app is waiting on an update.

PDFs, on 24 of the 59 models

• Models: 24 of 59, including four available without a subscription: GPT-5.4 Nano, Gemini 3.5 Flash Lite, Gemini 3.1 Flash Lite and Gemini 2.5 Flash Lite.

• Plan: any, free included.

• What reaches the model: usually the text we extract from it, with the original file kept for image-heavy pages and some scans.

• If the model cannot take it: "This model doesn't support PDF uploads. Switch to a model that accepts PDFs."

• If your plan cannot: not a plan limit.

• Verdict: choose the model before you choose the file, because 24 is the gate you will actually meet.

A PDF made of real text now goes to the model as the text we extract from it, whichever of the 24 you pick. That extraction finds nothing in a scan, so since 19 September a fully scanned PDF of up to 20 pages is transcribed first when you are signed in, and the model is told the transcription may contain recognition errors. Check any figure you act on against the page. A longer scan, one sent without an account, or one with a page too unclear to transcribe goes to the provider as the raw file, and what a provider makes of it is not something we would tell you to rely on. Switch that file to a model without native PDF support and it arrives as the line "[PDF attached: yourfile.pdf - content not available for this model]".

Text and code files, on all 59 models

• Models: all 59. GPT 4o, which is being retired and no longer counts as active, is the one model in the picker that will not take one.

• Plan: Pro and Pro+. A free account cannot attach one at all.

• What reaches the model: the file's text, with its filename.

• If the model cannot take it: "This model doesn't support text or code files."

• If your plan cannot: "Code file uploads are a Pro feature"

• Verdict: all but universal on the model side, so your plan is the only real question.

The server applies this independently of the picker. A free account's upload request comes back as a 402 carrying the same Pro message for anything that is neither an image nor a PDF, so there is no client behaviour to route around. The payoff of handling these as text is that a .md, .csv, .py or .json file lands in the conversation as plain text, which is a form every one of those 59 models can work with, and the filename travels with it.

Office and OpenDocument files, on all 59 models

• Models: all 59, the same set as text and code files, because these are turned into text before they leave us.

• Plan: Pro and Pro+.

• What reaches the model: text extracted on our servers. Never the original file.

• If the model cannot take it: "This model doesn't support document uploads."

• If your plan cannot: "Document uploads are a Pro feature"

• Verdict: dependable for prose, lossier than you would guess for spreadsheets.

A .docx, .xlsx, .pptx, .odt, .ods or .odp is a zip archive of XML, so we unzip it and pull the visible text in document order. The pre-2007 .doc, .xls and .ppt formats go down a separate path. Prose comes through this well, spreadsheets much less so, since the reader takes each cell's stored result and leaves the formulas and the number formatting behind. Where a file holds no readable text at all, a slide deck of images being the usual case, the extraction reports that nothing was found rather than passing on an empty document.

Video, on 6 of the 59 models

• Models: 6 of 59, every one of them Gemini. No model from any other provider in the catalogue takes video.

• Plan: Pro+ only.

• What reaches the model: the video file.

• If the model cannot take it: "This model doesn't support video. Switch to a Gemini model to send videos."

• If your plan cannot: "Video uploads are a Pro+ feature"

• Verdict: two narrow gates rather than one. Check your plan first.

Three of the six are on the no-subscription model list, and that changes nothing here, because the model gate and the plan gate are independent and a clip has to clear both. The order those checks run in changed on 12 September: the plan is now stated before the size. Until then a free account attaching a 40MB clip was told its plan supports files up to 10MB, which was true but left the actual blocker unsaid. The message now names Pro+. Web, mobile and the upload route all agree on that order, so if you still see a size message on a video, you are on a build made before the change.

Anything else, on none of the 59 models

• Models: none. Audio files, archives and unidentified binaries are refused before any model is consulted.

• Plan: no plan unlocks them.

• What reaches the model: nothing. The upload never starts.

• If the model cannot take it: "This file type isn't supported."

• If your plan cannot: not a plan limit.

• Verdict: convert it or paste its contents.

Audio is the one people reach for most. There is no audio attachment path on any plan, so a recording has to become text before it can be discussed. A file we turn away never reaches a server log, so the composer notes the extension it refused, and that tally is what ranks the formats worth adding next.

Limits by plan

Free is the only plan with a daily upload quota you would meet in ordinary use, and its three reset on a fixed 24-hour window aligned to UTC midnight rather than to your own, so the hour at which tomorrow's uploads arrive depends on where you are. Spend them and the attach button stops opening the picker, showing "You've used your 3 free uploads for today. Upgrade to Pro for more." instead. Pro is $20.99 a month or $199.99 a year, Pro+ is $34.99 a month or $349.99 a year, and Pro+ is the only plan that takes video. Current figures are on the plans section.

A file has to fit as well as attach

Passing every gate is not the same as fitting. A long PDF or a video eats into the same budget your conversation is drawing on, and the meter under the composer starts warning you when answer quality is likely to slip for the model you have chosen. Each plan also caps a single message, typed text and attachments together, at about 128,000 tokens on a free account, 200,000 on Pro and 256,000 on Pro+. Our explainer on what a context window is covers what that budget is and how to tell when you are near the end of it.

Where your upload ends up

An upload is a copy, and it leaves for whichever provider serves the model you picked. We do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the web picker. Most carry none, so an absent notice leaves you no wiser either way. Worth a moment's thought before you attach a payslip, a passport scan or a client contract: the gates in this article decide whether a file can be read, not who ends up holding it.