Nano Banana 2.1, Pro, GPT Image 2.5, Seedream or FLUX.2: 5 checks

Do not compare five image models on taste first. Ask five yes/no questions in a fixed order, and most of the field drops out on a hard limit before price matters. On Sume the order is: do you need a 1:8 or 8:1 shape, more than 10 reference images, a transparent background, a native 4K tier, and only then what you will pay per image.

Nano Banana 2.1 launched on Oct 6, and Google's model page (read 2026-10-12) lists 1K by default with 2K and 4K, wide ratios 1:4, 4:1, 1:8 and 8:1, and fusion of up to 14 reference images. A Sume request does not take everything Google's page does, so the checks below use what the Image API docs and the live catalog list.

The five rows side by side

These are the catalog values for the five ids: google/nano-banana-2.1, google/nano-banana-pro, openai/gpt-image-2.5, bytedance-seed/seedream-5-lite and black-forest-labs/flux.2-pro. Prices are what Sume bills per image, from each row's pricing line.

Check 1: do you need 1:8 or 8:1?

Of these five, only Nano Banana 2.1 lists 1:8, 8:1, 1:4 and 4:1. Nano Banana Pro stops at 21:9 and 9:16. FLUX.2 Pro reaches 21:9 and 9:21. GPT Image 2.5 and Seedream 5 Lite stay inside roughly 16:9. A request for an unlisted ratio gets 400 unsupported_parameter, so there is no silent crop to a nearby shape.

If your answer is yes, the decision is made and you can stop here.

Check 2: more than 10 reference images?

Google's page says 2.1 fuses up to 14 reference images. On Sume, Nano Banana 2.1 lists 10 in its input_references descriptor, as do Pro, Seedream 5 Lite and FLUX.2 Pro. GPT Image 2.5 lists 16. If you really need 11 to 16 sources in one call, it is the only one of the five that accepts them. Otherwise split the sources across two calls and feed the first output back in as a reference.

Check 3: transparent background or a mask?

background: transparent and mask_url are documented only for ChatGPT Image 2.5 (both the Flare and Sunburst ids). The other four rows do not list them, so they reject the field rather than ignore it. Pair them with a cutout step if you need alpha from a different model.

Check 4: native 4K?

Nano Banana 2.1 and Pro are the two rows with a resolution enum that includes 4K: $0.20 and $0.375 per image. GPT Image 2.5 can reach large sizes with custom image_size pixels (both edges multiples of 16, maximum edge 3840, aspect ratio at most 3:1), but the price then follows image tokens rather than a flat tier. Seedream 5 Lite and FLUX.2 Pro list no tier field.

Check 5: price, once the limits agree

If two or more rows survive, the price ladder from the table settles it: FLUX.2 Pro at $0.0375, Seedream 5 Lite at $0.04375, GPT Image 2.5 at $0.065875 for a high-quality 1024 square, Nano Banana 2.1 at $0.10 and Pro at $0.1875. For 1,000 images at the listed default that is $37.50, $43.75, $65.88, $100.00 and $187.50. Pro bills the same $0.1875 at 1K and 2K, so if you pick Pro, use 2K.

• Flat per-image rows (2.1, Pro, Seedream, FLUX.2 Pro) cost the same with one reference or ten.
• GPT Image 2.5 is billed from image tokens, so a high-quality or reference-heavy call can cost more than the catalog's 1024 figure.
• Failed generations are not billed on any row.

A note on the 30-second wait

POST /v1/images waits up to 30 seconds and returns 200 with the image. The docs say 4K, high quality and large n are the likeliest to fall back to a 202 job envelope, which you then read from GET /v1/jobs/{id}/result. Branch on the status code, not the body shape, whichever row you pick.