31 lookbook edits, 3 reference images each: 93 inputs, flat rows

The short answer

Thirty-one lookbook edits at 2K, each sent with three reference images (a garment, a model photo, a background), cost $4.65 on Nano Banana 2.1 and $5.81 on Nano Banana Pro. That is 93 reference inputs in total, and they do not appear in the arithmetic: the catalog row is per output image.

GPT Image 2.5 is the case to watch. Its upstream rates are per token, so confirm what a reference-heavy call costs before you scale it.

How to send the references

On POST /v1/images, put the references in input_references as image_url objects with public HTTPS URLs. Sume rejects localhost, private-network and non-HTTPS URLs before submission. A model whose input_references descriptor is min 0, max 0 is text-to-image only and rejects references, so read the descriptor from GET /v1/images/models first.

For an edit, send aspect_ratio set to auto to match the reference. Omitting the field is not the same as auto, and the docs say so directly.

The arithmetic

Cost of the run is cost_usd x n per call, and n is 1 here. 31 x $0.15 = $4.65; 31 x $0.1875 = $5.81. Pro at 2K costs $1.16 more across the run, or $0.0375 per edit.

For GPT Image 2.5 the table shows the catalog's medium 2K row, $0.055625, which is a default-options estimate.

What to verify before you scale

Run one edit per model with your real references and read usage.cost in the response. The docs describe GPT Image 2.5 as token-billed upstream with input image tokens at their own rate, so a call carrying large references can differ from the default-options row. The banana rows are what the catalog publishes per image.

Keep the three references to what the edit needs. Extra references do not change the banana price, but they do make the instruction harder for the model to follow, and that costs you retries.

Choosing between the two banana models

The gap between the models is $1.16 over 31 edits, which is small against the cost of a bad edit that you have to redo. If 2.1 needs 25 percent more attempts per accepted image than Pro, the two cost the same per accepted image, because a retry costs a full extra row.

That is a threshold to test, not a result to assume. Run six edits on each model with the same three references, count the accepted ones, and divide the spend by the accepted count. The docs give you the price side of that ratio; only your own images give you the other side.

Name each image's role in the prompt (garment, model, background) instead of relying on the order of the array, since the docs describe input_references only as an array of image URLs.