9 low GPT Image 2.5 drafts cost about the same as 1 high 4K image

Nine low-quality 1K drafts from GPT Image 2.5 on Sume cost $0.22275 together, which is $0.00025 more than a single high-quality 4K image at $0.2225. Four medium 2K images cost 4 x $0.055625 = $0.2225, exactly the price of that one 4K image. The tiers are close to a clean 1 : 2.25 : 9 ladder, so you can plan drafts and finals in units of the cheapest tier.

This is a planning shortcut, not a promise about every request. The three prices below are the Sume list prices for default options, and they bundle quality and size together: low is quoted at 1K, medium at 2K, and high at 4K. A low image at 4K or a high image at 1K is a different line item.

The three GPT Image 2.5 price points

Sume bills provider list price times 1.25. The figures in this table come from the lead-verified price list and match the GPT Image 2.5 catalog entry. The catalog formula ends in a round up to the cent, so a billed line can differ from the unrounded figure by a fraction of a cent. Edit calls with reference images also add input tokens, which the table does not include.

What nine drafts buy you

The useful reading is a budget rule. If you draft a layout at low quality and pick one winner to render at high 4K, you can afford eight drafts plus the final for $0.42050, or nine drafts and no final for $0.22275. With four medium 2K tries, you spend the same as one high 4K render and you get four chances to see the composition at a size where small text is readable.

Which tier you draft in depends on what you are testing. Layout, subject placement and color usually survive at low quality. Dense text, packaging copy and fine product detail often need medium or high, and the Sume Image 1.0 docs say to use a higher quality value for finals, dense text, or packaging.

• Layout and mood check: nine low 1K drafts, $0.22275
• Text legibility check: four medium 2K drafts, $0.2225
• One final render: one high 4K image, $0.2225

Set quality and size on purpose

The Image API documents that if you omit quality for GPT Image 2.5, the default is high, and that auto quality reserves the max tier when Sume admits the job. A script that forgets the field therefore spends the most per image. Pass quality explicitly in every draft call.

For custom pixels, both edges must be multiples of 16, the longest edge is at most 3840, the aspect ratio is at most 3:1, and the image must have between 655,360 and 8,294,400 pixels. A 3840 x 2160 final sits exactly on the upper pixel limit. The named presets and the custom size field are the way to pin the size tier you priced.

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Poster layout: bold headline top left, product bottom right",
  "quality": "low",
  "image_size": "1024x1024"
}

When the arithmetic breaks

Three things move the numbers. Reference images on an edit add estimated input tokens to the admission hold. Quality values above high (xhigh and max) cost more than high at the same size. And a model switch changes everything: a flat-priced row such as Seedream 5 Lite at $0.04375 per image does not scale with quality at all. Compare per-image prices for the same task before you commit a batch, and read usage.cost in each response to confirm what Sume billed.