gpt-image-1 low, medium, high on GPT Image 2.5: Sume prices

If your gpt-image-1 code sends quality: low, medium or high, the same three words are valid on GPT Image 2.5, and on Sume they cost $0.0074, $0.0165 and $0.0659 per 1024x1024 image. The new model adds two tiers above high: xhigh at $0.1171 and max at $0.2635. If you omit quality, Sume sends high, so a bare call costs $0.0659.

The reason to look now is the calendar. OpenAI's deprecations page (read 2026-10-05) lists gpt-image-1 as shutting down on 2026-10-23 and names gpt-image-2.5-sunburst or gpt-image-2.5-flare as its replacement. This post covers only the Sume side of that move: the six quality values and what each one bills. It makes no claim about gpt-image-1's own prices, because the deprecations page does not list them.

Price by quality at 1024x1024

Sume bills image generation per image from Fal's token rates ($30 per million output image tokens) times 1.25, and the Image API page states the 1024x1024 output rates for xhigh and max. The table adds the other tiers from the same token formula. Flare (openai/gpt-image-2.5) and Sunburst (openai/gpt-image-2.5-sunburst) quote the same amounts.

What the quote includes

The Fal column is the provider rate before Sume's 1.25 factor, and Fal rounds a total up to $0.0001, so a quote can sit a hair above the raw token arithmetic. Prompt text is extra: at $5 per million input text tokens, a 700-token prompt adds about $0.0035 before the factor, which is a rounding error on a high image and a visible share of a low one. If you generate thousands of low-quality drafts, keep the prompt short.

Edits add input image tokens at $8 per million on top of this, which is why the same tier costs more when you send a reference. The size is fixed at 1024x1024 here so the tiers compare cleanly; other sizes shift every row by the same pattern.

What changes in your request

Only the model string and the endpoint differ. A minimal call looks like this, with quality pinned so the price is known before you send it:

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": "A ceramic mug on a pale oak table, soft window light",
    "quality": "medium",
    "image_size": "1024x1024"
  }'

Three rules for the migration

These three behaviors decide most surprises on the invoice, and each is stated on the Image API page.

• Omitted quality means high on Sume, so leaving it out is a 4x decision compared with medium. Set it on every call.
• auto is accepted, but it reserves the max price while the job runs. Use it only when you can afford the hold; use an explicit tier otherwise.
• xhigh and max exist only on ChatGPT Image 2.5. The older openai/gpt-image-2 row on Sume lists low, medium and high, so code that sends xhigh to it fails with a 400.

How to choose a tier

Treat the tiers as a price ladder, not a quality promise: each step up costs a fixed multiple, and high is 8.9x the price of low. A workable habit is to run the whole prompt set at low first, look at the composition, and re-run only the keepers at high. Ten drafts at low cost $0.074, which is less than one high image at $0.066.

Sizes move the numbers too. The table above is square; a 1536x1024 image at high quotes lower, which the size breakdown covers.

Before you cut over

Send one call per tier from staging and compare usage.cost in each response with the table. Sume reports the billed USD amount in that field, and a generation that fails is not billed. If a call returns 202 with a job envelope instead of 200, the image is still rendering; fetch it from the job result.