Which Sume image model for what job: one price table, Oct 2026

For most jobs on Sume, start with ChatGPT Image 2.5 (openai/gpt-image-2.5) at 1 to 7 cents, and move to a different row only for a reason you can name: Nano Banana Pro or 2.1 for photoreal edits and extreme ratios, Ideogram 4.5 for dense text, Seedream or Qwen Image for cheap bulk drafts. The table below maps each job to a row with its catalog id and billed price.

The table

Prices are the catalog list times 1.25, rounded up to the cent, at each row's default tier. GPT Image 2.5 shows its low, medium and high tiers. Ideogram 4.5 shows low, medium and high.

How to read it

Cost per image differs by a factor of nearly 30 between the cheapest row and ChatGPT Image 2, which is 27 cents. But cost per usable image matters more. A row that needs four tries at 3 cents costs 12 cents per keeper, more than a 7-cent row that works on the first.

Price is not the only constraint. Text-only rows (Imagen 4 Fast and Ultra, Recraft V4, Qwen Image Max) do not accept references, so they cannot do an edit from your photo. Ratio lists differ too: only Nano Banana 2.1 lists 4:1 and 8:1, and GPT Image 2.5 does not list 3:2.

Let Sume choose, or pin

model: "sume/auto" lets Image Router choose. The docs say Sume never discloses which family ran and that Auto routing continues to use Flare for ChatGPT Image 2.5. That is fine for exploration. For brand work or a comparison, pin an id, because Auto may pick differently next week.

Whatever you choose, read GET /v1/images/models once at the start of a project and save the response. Rows come and go: Sume says Image 1.0 is retiring, and Nano Banana 2 requests already run as Nano Banana 2.1.

A simple decision order

Ask these questions in order and stop at the first that applies.

• Need exact words in the picture: try Ideogram 4.5 or GPT Image 2.5 first.
• Starting from a photo: pick a row with edit support; check the references limit (16 for GPT Image 2.5, 5 for Ideogram 4.5).
• Need a banner ratio like 4:1: Nano Banana 2.1.
• Need hundreds of drafts: Qwen Image, Seedream or GPT Image 2.5 at low.
• Need a transparent background: GPT Image 2.5 with background: "transparent".

Keeping the table honest

Prices drift and rows change. Rebuild this table from GET /v1/images/models and the per-model endpoint records whenever you start a project, and note the date next to it. A row that is cheapest today may be retired next month, and the endpoint pricing lines are the amounts Sume actually charges to your wallet.