Same image, each vision model
How each model tokenises your image, and the monthly input cost at your volume — ranked cheapest first. Prompt tokens per image are included in the monthly figure.
| Model | Image tokens | Per image | Per month |
|---|
An image is input tokens in disguise
When you send a picture to a vision model, it doesn't read pixels for free — it converts the image into tokens and charges them at the input rate, right alongside your text prompt. The catch is that every provider counts those tokens differently, and none of them count by file size. OpenAI's GPT-4o fits the image into a grid of 512-pixel tiles and charges 85 base tokens plus 170 per tile. Anthropic's Claude uses roughly width times height divided by 750. Google's Gemini charges a flat 258 tokens for small images and tiles larger ones. The same 1024×1024 screenshot can therefore be 765 tokens on one model and 1,400 on another — a near-2× swing before you've written a single word of instruction.
That matters most at volume. A document-processing pipeline, a moderation feed or a "chat with your screenshots" feature can push tens of thousands of images a day, and at those counts the image tokens dwarf the prompt. This calculator makes the trade-off visible: it shows how each model tokenises your exact resolution and what that costs per month, so you can pick the cheapest vision model for your image mix and see instantly how much you'd save by down-scaling. Pair it with the token counter for the text side and the cheapest LLM API tool for the overall model choice.
How to use it
1. Enter the width and height of the images you send — the resolution after any resizing you already do.
2. Pick OpenAI's detail mode: high reads fine text and small labels, low flattens to an 85-token thumbnail.
3. Enter how many images you send per month, and optionally the prompt tokens that ride along with each.
4. Read the ranked table to see the cheapest vision model for your image size and the monthly cost.
Common mistakes
Sending full-resolution when a crop would do. Token cost grows with pixel area, so a 2048² image can cost four times a 1024² one for detail the model never needed. Leaving detail on high by default. If you only need the gist, low-detail mode is a flat 85 tokens — often a 5–10× saving. Forgetting the prompt still counts. The image tokens are on top of your text prompt and any system prompt, so the real per-call cost is both added together. Ignoring output. This tool prices the image as input; the model's written answer is billed separately at the output rate.
FAQ
Does a bigger file cost more?
No — token cost tracks pixel dimensions, not bytes. A heavily compressed 2048² JPEG and a crisp 2048² PNG cost the same number of tokens; only the width and height matter.
How do I cut image input cost the most?
Down-scale to the smallest resolution that still shows the detail you need, and use low-detail mode wherever the task only needs the gist. Together those two levers routinely cut vision input cost several-fold.
Are these the exact provider formulas?
They're faithful reference implementations of the published 2026 tile and pixel rules, but providers round and revise them. Treat the token counts as tight estimates and confirm against your provider's own tokeniser for billing-grade numbers.
What about video?
Video is typically charged as a sequence of sampled frames, each priced like an image. Estimate the frames-per-second the model samples, multiply by seconds, and treat each frame as one image in this tool.
Estimate only. Vision tokenisation and per-million input rates change — verify with your provider before relying on these figures.