GPT Image 2.5 on Sume: each reference adds about 27% to the quote

Each reference image you add to a GPT Image 2.5 request raises the Sume quote by about 27% of the output price: a medium 1024x1024 call goes from $0.0165 with no reference to $0.0209 with one and $0.0867 with sixteen. The step is $0.0044 for the first reference at that tier, and the later ones add about the same.

The reason is in Sume's estimator. Sume's Image API page says input image tokens are billed at $8 per million, against $30 per million for output image tokens, and that the counts are estimates. The estimator assumes one reference is about as many tokens as the output image, so each reference costs 8/30 of the output price, or 26.7%, before Sume's 1.25 factor.

Price by reference count

All rows are medium quality at 1024x1024 with a short prompt. The second column is Sume's quote, and the third is the change from the text-to-image price.

The same rule at other tiers

Because the reference cost is a share of the output cost, the percentage is about the same at every tier and the dollars scale with the tier. One reference at low takes the quote from $0.0074 to $0.0094, a rise of 27%, and at high from $0.0659 to $0.0835, also 27%. Three references at low cost $0.0132, which is 80% more than text-to-image.

That makes references a larger share of the bill at the low tiers. At low three references nearly double the price. At high three references add the same 80%, but the dollars are larger: $0.0527 more per call.

How to keep the cost down

Reference images are often what makes an edit work, so cut them only where they do not.

• Send only the references the instruction uses. Sixteen is the limit, but a call with four costs $0.0341 at medium, less than half of the 16-reference price.
• Draft with fewer references at low, then add the extra ones for the final render.
• Use mask_url to restrict an edit, not extra references that describe the same region.
• Check usage.cost on the first response of a new workflow against this table, because the input counts are estimates.

Planning a batch

For a batch of 1,000 edits at medium with two references each, the quote is $25.25, against $16.50 for the same 1,000 images with no reference. The references account for $8.75 of that. If the same two references are reused on every call, the cost still repeats on every call, because each request carries its own input tokens.

This also changes the decision between one call with n: 4 and four separate calls. Four images in one call are priced as four images, and the references are counted once per output image in the estimator, so a single call with four outputs does not make the references free. Compare a real usage.cost before you assume a saving, and do it once per workflow.

Finally, treat the 27% as a rule of thumb for budgeting, not as a rate card. The estimator is allowed to change its input-token assumption, and the live usage.cost and the catalog are the sources of record, and a quick re-measure each quarter is cheap. Record the date and the figure next to the number in your own budget sheet so a later reader knows how old it is.

Request

Two references in image_urls, medium quality. The response should quote close to $0.0253.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Put the jacket from the first image on the person in the second",
    "image_urls": [
      "https://example.com/jacket.png",
      "https://example.com/person.png"
    ],
    "quality": "medium",
    "image_size": "1024x1024"
  }'