What an image actually costs, and why it moved when you changed model

The first surprise is that there's no image line to look for. Providers convert a picture into input tokens and charge the ordinary input rate, so a vision call looks like a text call with a large prompt. Anthropic publishes the conversion: an image is divided into 28-pixel patches and each patch is one visual token, so the count is the width in patches times the height in patches. That makes image cost a function of resolution, which is the second surprise, because resolution is usually set by whatever produced the file rather than by anyone thinking about money. The third is the one that moves bills. Models sit in one of two resolution tiers, and the newer tier permits roughly three times the visual tokens for the same image before downscaling kicks in. Upgrade the model and your image costs rise sharply with no change to your code, your prompts or your pictures.

• Your bill rose after a model upgrade and the prompts didn't change.
• You looked for a per-image charge and couldn't find one.
• Screenshots go to the model at whatever size the screen happened to be.
• Nobody knows the average resolution of the images you send.
• A vision feature costs more than a text feature doing similar work.

1. Take a typical image you send and work out its visual tokens from its dimensions.
2. Multiply by your model's input rate and by your monthly image volume.
3. Check which resolution tier your model sits in, and what the other tier would cost.
4. Find the largest image your product will accept and price that as your worst case.

Anthropic's vision documentation, read 2026-08-05, publishes the formula, a token table and worked per-thousand figures. A 1000x1000 image costs ceil(1000/28) x ceil(1000/28) visual tokens. A 4K image at 3840x2160 downscales to 1,560 visual tokens on the standard tier and 4,784 on the high-resolution tier, which covers Claude 4.7 and later. Rates are the price book's real $1.00 per million for Haiku 4.5 and $5.00 for Opus 5. Every figure below is published rather than assumed.

Anthropic states all three dollar figures in its own prose, and re-deriving them from the formula reproduces each one exactly. That's worth noticing: the arithmetic here isn't an interpretation of the pricing page, it agrees with it to the cent. The 3.07x tier difference is the part that arrives unannounced, because a model upgrade changes it without touching a line of your code.

The formula is public and the rates are public, so anyone can price one image. What nobody outside your system knows is the distribution of images you actually send, and that's the whole number. A product accepting user uploads has a resolution profile it has never looked at, with a tail of very large files that cost multiples of the median. Culpa prices every call from the tokens the provider reports, so image cost is already inside your per-call figures rather than sitting in a separate bucket, and it carries whatever identifiers you attach. That turns the useful questions into queries: which feature sends the biggest images, which customers, and what a downscaling step would have saved on last month's real traffic rather than on an assumed average.

• Thinking Token Cost: You Pay for Reasoning You Never See
• Tool Definition Token Cost: You Pay for Every Unused Tool
• What is token accounting
• Input Token Creep: Four Percent a Month Is 60% a Year