2026-10-03

Cheapest Sume image models that take a reference photo

Reference edits start at $0.025 per image on Sume. Ranked list of every catalog model that accepts input_references, with reference limits and prices.

5 min readSume

All posts

The cheapest reference edits on Sume are Qwen Image and Grok Imagine at $0.025 per image, then Seedream 4 at $0.0325 and Flux 2 Pro at $0.0375 (catalog read 2026-10-03). GPT Image 2.5 takes 16 references, the most of any row, and its medium square estimate is $0.0165, which is cheaper than any of them at that quality. Five rows are text-only: Imagen 4 Fast and Ultra, Recraft V4, Qwen Image Max and Higgsfield Soul accept no references.

Ranked list

The list comes from the catalog descriptors and the per-image endpoint prices (list x 1.25), read 2026-10-03. GPT Image 2.5 prices are repo estimates by quality and size, not billed figures.

Reference-capable models by price (read 2026-10-03)
Rank	Model	Price per image	Max references
1	Qwen Image	$0.025	10
1	Grok Imagine	$0.025	10
3	Seedream 4	$0.0325	10
4	Flux 2 Pro	$0.0375	10
4	Ideogram 4.5, low quality	$0.0375	5
6	Seedream 5 Lite	$0.04375	10
7	Seedream 4.5	$0.05	10
8	Flux 2 Flex	$0.0625	10
9	GPT Image 2.5, high (default estimate)	$0.0659	16
10	Ideogram V3	$0.075	10
10	Ideogram 4.5, medium (default)	$0.075	5
12	Nano Banana 2	$0.10	10
13	Nano Banana Pro	$0.1875	10

Where GPT Image 2.5 fits

GPT Image 2.5 is the only row where quality sets the price. A square at low quality is an estimated $0.0074, medium $0.0165, high $0.0659, xhigh $0.1171 and max $0.2634. It also takes background and mask_url, so it is the row for masked edits and transparent output. The default quality is high, which is why a single catalog figure of $0.0659 shows.

Edit request

Send the photo as input_references and ask for one change.

curl -s https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen/qwen-image", "input_references": [{"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}}], "prompt": "Keep the person and pose, change the background to a sunny balcony"}'

What to test first

Price tells you little about edit fidelity. Run your three hardest pictures through the top three rows and compare the results by eye.

Does the face or product stay the same?
Does it follow the one change you asked for?
How many takes did you need?

Sources

Image API