Costs & Limits of Vision

VisionCost

Part of: Multimodal AI: Vision & Images

A founder wired a vision model into a "snap a photo of your fridge" app, shipped it, and watched the bill explode. The culprit was not traffic. Every user was uploading 12-megapixel phone photos, and each one quietly became thousands of image tokens. The app worked perfectly and lost money on every scan. Vision is powerful, but pixels are not free, and knowing the cost math is what separates a demo from a product. What it is Vision cost is the extra token spend an image adds to a request. An image isn't billed as "one image." It is converted to image tokens : recall the patches from lesson 1, and those tokens are billed like any input tokens. The higher the resolution, the more patches, the more tokens, the bigger the bill. Most APIs expose a detail level (often "low" and "high"). Low detail downscales the image to a small fixed size, capping its token count cheaply. High detail keeps more resolution and costs far more. How it works The cost of a vision request is text tokens plus image tokens plus the output: Three levers control the bill: 1. Resolution. Token count grows with area, so doubling each side roughly quadruples image tokens. 2. Detail level. "Low" caps tokens for coars

Challenge: The Detail Router