What one image costs to analyse β resolution becomes tiles becomes tokens.
Vision billing is really a resolution tax: every 512px tile adds tokens. Resize before upload and the per-image cost drops fast. For creating images rather than reading them, use the image generation cost calculator, and blend input/output rates on the blended LLM price calculator.
The Vision API Cost Calculator estimates what you'll pay to send images to a GPT-4o-style vision model. You enter an image's width and height in pixels, and the tool converts those dimensions into the number of input tokens the model charges for, since vision models tile and price images by resolution rather than by file size. Multiply that token count by your $ per 1M input tokens rate to get a per-image cost, then by your images per month to project a monthly bill. The two biggest drivers are image resolution, which sets tokens per image, and monthly volume, which scales the total.
The key trade-off to watch is resolution versus cost: larger images consume more tokens, so a small dimension increase can meaningfully raise your bill at scale. Before sending, downscale images to the smallest size your task still reads accurately, and estimate high-volume workloads here first so a pricing surprise doesn't land on your invoice.
Vision models split an image into 512-pixel tiles. GPT-4o-style pricing charges a fixed base (about 85 tokens) plus roughly 170 tokens per tile, then bills those as normal input tokens. A 1024Γ1024 image is four tiles, so about 765 tokens β the more megapixels, the more tiles and tokens.
Downscale images before sending them β resolution drives tile count, and most tasks do not need full resolution. Use low-detail mode when the model offers it (a flat, cheap token cost), crop to the region that matters, and avoid re-sending the same image across turns of a conversation.
Cost at different volumes:
| Volume | Monthly | Yearly |
|---|