Every vision API has a size rule, and none of them are the same rule. Claude caps an image at 8000x8000 px and 10 MB base64, bills 28x28-pixel patches as visual tokens, and tightens the per-image dimension limit the moment a request carries more than 20 images. OpenAI accepts PNG, JPEG, WEBP and non-animated GIF up to a 512 MB total payload with 1,500 images per request, but rejects any single image that needs more than 30,000 patches after resizing instead of resizing it down for you. Gemini takes PNG, JPEG, WEBP, HEIC and HEIF, up to 3,600 files per request, and lets media_resolution decide how many tokens each image can spend. The shared consequence: a 12-megapixel phone photo is mostly wasted bytes. It gets downscaled server-side anyway, so you pay for tokens or payload you never use, and in the OpenAI case you can hit a hard rejection instead of a graceful downscale. Reducing the file before the request puts you back in control of what the model actually sees. Skwish resizes to an exact dimension, converts format and compresses in the browser with a per-file before and after readout, and batches a whole folder into one ZIP.