2026-10-10

Sume image catalog on 2026-10-10: 19 ids from 0.5 to 26 cents

All 19 Sume image model ids sorted by base price, with how many take references, lists auto, offer a quality field, a resolution field, or a mask.

5 min readSume

All posts

As of 2026-10-10 the Sume image catalog lists 19 model ids, priced from 0.005 USD (Higgsfield Soul) to 0.26375 USD (GPT Image 2) per image at base price. Fourteen of them accept reference images, six list auto as an aspect ratio, five have a quality field, five have a resolution field, and only the two ChatGPT Image 2.5 ids take a mask_url.

I read these from the capability descriptors and endpoint pricing in the catalog code; the live source for any decision is always GET /v1/images/models.

The ladder, cheapest first

Prices are the base price per image on the sume endpoint. A request can cost more when you raise quality or resolution on a model that lists those fields, so check pricing for your exact settings.

Sume image catalog by base price per image (read 2026-10-10)
Model id	Base price (USD)	References	Extra fields
higgsfield/soul	0.005	0	resolution 720p/1080p; n 1 or 4
x-ai/grok-image	0.025	10	n fixed at 1
qwen/qwen-image	0.025	10	none
google/imagen-4-fast	0.025	0	none
bytedance-seed/seedream-4	0.0325	10	auto ratio
black-forest-labs/flux.2-pro	0.0375	10	none
bytedance-seed/seedream-5-lite	0.04375	10	none
bytedance-seed/seedream-4.5	0.05	10	none
recraft/recraft-v4	0.05	0	webp output only
black-forest-labs/flux.2-flex	0.0625	10	none
openai/gpt-image-2.5 and -sunburst	0.065875	16	quality (6), background, mask_url, auto ratio
ideogram/ideogram-v3	0.075	10	quality (3)
ideogram/ideogram-v4.5	0.075	5	quality (3), resolution 1K/2K
google/imagen-4-ultra	0.075	0	resolution 1K/2K
qwen/qwen-image-max	0.09375	0	none
google/nano-banana-2.1	0.10	10	resolution 512-4K, auto ratio
google/nano-banana-pro	0.1875	10	resolution 512-4K, auto ratio
openai/gpt-image-2	0.26375	10	quality (3), auto ratio

What the ladder shows

Price does not track capability in a straight line. The cheapest reference-capable ids, Grok Image and Qwen Image, cost 0.025 USD; Qwen Image Max costs 0.09375 USD, 3.75 times as much, and takes no references at all. GPT Image 2.5 at 0.065875 USD is cheaper than GPT Image 2 at 0.26375 USD and accepts 16 references against 10.

So the question is rarely which model is best. It is which fields your request needs: references, a ratio, a mask, a tier. Filter by those first, then sort what is left by price.

The five edit-incapable ids

Five ids are text-to-image only, with an input_references range of 0 to 0: Higgsfield Soul, Imagen 4 Fast, Imagen 4 Ultra, Recraft V4 and Qwen Image Max. A request that sends a reference to any of them returns 400 unsupported_parameter. If your pipeline edits photos, remove these from the candidate list in code, not by hand.

The other fourteen take references, but the ceiling varies from 5 on Ideogram 4.5 to 16 on ChatGPT Image 2.5, so a request with eight references works on most and fails on Ideogram.

Using the table

Treat this as a snapshot. Models are added and retired, and Sume documents that retired ids can keep working as aliases: Nano Banana 2 still runs, as Nano Banana 2.1. Read the catalog at start-up, pin the ids you use, and compare against a saved copy so that a change is an alert and not a surprise.

Need references and the shape of the source: Seedream 4, Nano Banana 2.1 or Pro, GPT Image 2 or 2.5.
Need a mask: GPT Image 2.5 or its Sunburst id.
Need the lowest cost for text-only drafts: Soul, then Imagen 4 Fast.
Need a tall or wide strip: check the ratio list before the price.

Sources

Image API