If you snap a photo on your phone and drop it straight into ChatGPT, Claude or Gemini, you are usually uploading far more pixels than the model will ever look at. The assistant quietly shrinks the image first, and you have no say in how. The practical answer for most people: resize photos to about 1536 pixels on the long side and save them as JPEG. For screenshots full of small text, crop to the part that matters instead of shrinking the whole thing.
Below are the limits each vendor publishes, a worked example with a real 12-megapixel phone photo, and the edge cases where that rule of thumb does not hold. Every limit in this article comes from the vendor's own documentation or help center, checked on October 19, 2026; links are at the end.
Quick answer by use case
| What you are uploading | Target size | Format |
|---|---|---|
| Everyday photo (a plant, a spotted leaf, a receipt on a table) | 1536 px long edge, e.g. 1536 x 1152 for a 4:3 photo | JPEG |
| Screenshot of an app, spreadsheet or error message | Crop to the relevant area, keep native pixels, ideally under 1568 px on the long edge | PNG |
| Photo of a printed page or form you want read word for word | Crop to the page edges, 1500 to 2000 px long edge | JPEG or PNG |
| More than 20 images in one Claude API request | 2000 px or less on both sides (required) | JPEG |
What each assistant officially accepts
ChatGPT
OpenAI's File Uploads FAQ sets a limit of 20 MB per image in ChatGPT, inside a general 512 MB per-file cap. The developer documentation lists PNG, JPEG, WEBP and non-animated GIF as supported image types, and says that all resizing "preserves aspect ratio without enlarging smaller images." In other words, uploading a bigger file never adds detail beyond what the model's own sizing rules keep, and upscaling a small image before uploading does nothing useful.
Claude
Anthropic publishes the most detailed numbers. Claude accepts JPEG, PNG, GIF and WebP (animated GIFs use only the first frame). The maximum is 8000 x 8000 px per image and 10 MB per image on claude.ai and the Claude API. Most models then downscale anything larger than 1568 px on the long edge or 1568 visual tokens. Claude 4.7 and later models use a high-resolution tier of 2576 px or 4784 visual tokens.
Claude splits an image into 28 x 28 pixel patches, and each patch is one visual token, so the cost is ceil(width / 28) x ceil(height / 28). That formula lets you predict exactly what happens to your file.
Gemini
Google's Gemini Apps help page allows up to 10 files per prompt and up to 100 MB for non-video files. For developers, Google's token documentation says images that are 384 px or smaller on both sides count as 258 tokens, and larger images are tiled into 768 x 768 pixel tiles at 258 tokens each.
Worked example: a 12-megapixel phone photo
A standard 12 MP shot from an iPhone or Pixel is 4032 x 3024 pixels, which is 12.19 million pixels. Here is what that photo costs compared with the same photo resized to 1536 x 1152.
| Image | Pixels | Claude native patch count | Claude standard tier | Gemini tile rule |
|---|---|---|---|---|
| Original 4032 x 3024 | 12.19 MP | 144 x 108 = 15,552 | Downscaled to about 1269 x 952, 1,564 tokens | 6 x 4 tiles = 6,192 tokens |
| Resized 1536 x 1152 | 1.77 MP | 55 x 42 = 2,310 | Downscaled to about 1269 x 952, 1,564 tokens | 2 x 2 tiles = 1,032 tokens |
| Resized 1024 x 768 | 0.79 MP | 37 x 28 = 1,036 | Not resized, 1,036 tokens | 2 x 1 tiles = 516 tokens |
Three things stand out:
- The resized photo has 85.5% fewer pixels (1.77 MP instead of 12.19 MP), so it uploads much faster on mobile data and sits comfortably under every file-size limit above.
- On Claude's standard tier, both versions end up at the same size. Anthropic's own table shows a 2000 x 1500 image shrinking to 1269 x 952, and any 4:3 photo larger than that lands in the same place. The extra 10 MP in the original is simply thrown away.
- On Claude 4.7 and later, the 1536 px version is kept at full size (2,310 tokens, well under the 4784 cap), while the original is cut down to roughly 2210 x 1658 by our calculation from the patch formula. Unless you need fine detail, 1536 px gives the model a sharp image at about half the token cost.
Under the tiling rule in Google's token documentation, the size you pick matters in steps: 1536 px is exactly two 768 px tiles wide, but going to 1568 px pushes the image into a third column of tiles (6 tiles, 1,548 tokens) for 32 extra pixels. Multiples of 768 are the efficient choices if you are paying per token through the Gemini API.
How to do it in under a minute
- Open the Toolspedia Image Resizer and upload the photo.
- Set the width to 1536 for a landscape photo (or the height to 1536 for a portrait one) and keep the aspect ratio locked, so a 4032 x 3024 photo becomes 1536 x 1152.
- Download the result as JPEG.
- If the file is still larger than you want for a slow connection, run it through the Image Compressor. A JPEG at this size normally needs only light compression.
- Upload it to the chatbot as usual.
Screenshots and documents: crop, do not shrink
Photos survive downscaling well. Screenshots do not, because the important detail is small text. Take a full-screen screenshot from a MacBook with a Retina display, 2880 x 1800 pixels. On Claude's standard tier, the formula says it gets cut to roughly 1389 x 868, which is less than half its width. Text that was 12 pixels tall becomes about 6 pixels tall, and that is where misread numbers and invented cell values come from.
The fix is to throw away the parts you do not need. Crop the screenshot to the error dialog or the block of spreadsheet cells you are asking about, for example a 1400 x 800 region. That works out to 50 x 29 = 1,450 tokens, under the 1568 limit, so Claude reads it at full sharpness with no resizing at all. The Image Crop Tool does this in the browser. Save screenshots as PNG: JPEG compression adds blurry halos around sharp text edges.
For multi-page documents, a PDF is usually a better upload than a stack of phone photos. Our guide on compressing a PDF before uploading it to ChatGPT, Claude or Gemini covers the file limits for that route.
Edge cases worth knowing
- HEIC photos from iPhones. None of the developer format lists above include HEIC. If an upload is rejected or a third-party app built on these APIs refuses the file, export it as JPEG first (on an iPhone, Settings, Camera, Formats, Most Compatible makes new photos JPEG).
- Lots of images at once. Claude's API rejects requests with more than 20 images unless each one is 2000 px or less on both sides. claude.ai allows up to 20 images per message.
- Tiny images. Upscaling a 400 px thumbnail to 1536 px adds no information. Send the original and, if possible, find a larger source.
- Very tall images. A full-page phone screenshot of 1179 x 2556 is limited by its height. On Claude's standard tier it shrinks to about 723 x 1568, so text gets smaller. Split it into two crops instead.
- Need exact detail? For counting small objects, reading serial numbers or checking fine print, skip resizing, crop tightly and send the crop at native resolution.
- File size targets. If a form or app asks for a specific size such as 100 KB, see how to compress an image to 100KB; that is a separate problem from what the AI needs.
Resize it yourself or let the chatbot do it?
Letting the assistant downscale is fine for a one-off question on a fast connection: you lose nothing on Claude's standard tier, as the worked example shows. Resizing yourself is worth it when you are on mobile data, when the file is close to the 10 MB (Claude) or 20 MB (ChatGPT) cap, when you are paying per token through an API, or when you want to choose what gets cropped instead of letting a whole screenshot be shrunk.
As for tools: Windows Photos has a Resize image option, and Preview on a Mac has Tools, Adjust Size. Both work well for one file on a computer. A browser tool like the Toolspedia resizer is handier on a phone or a shared work laptop where you cannot install anything. Batch scripts with ImageMagick or Python make sense only if you prepare dozens of images for an API pipeline.