8:1 strips and 21:9 banners: which image models take them

The short answer

For an 8:1 strip, Nano Banana 2.1 is the only one of the three models in this post that lists it. Nano Banana Pro does not list 8:1, 4:1, 1:4 or 1:8, and GPT Image 2.5 caps the aspect ratio at 3:1 for custom pixels, so 8:1 is out of its range. For 21:9, all three can do it, and the price depends on the model and tier, not on the ratio.

Sume rejects a parameter value a model does not list. Check the catalog before you pin a ratio.

What each model lists

Nano Banana Pro lists auto, 21:9, 16:9, 3:2, 4:3, 5:4, 1:1, 4:5, 3:4, 2:3 and 9:16. Nano Banana 2.1 lists those plus 4:1, 1:4, 8:1 and 1:8. The docs say a model accepts only the values its catalog descriptors list.

GPT Image 2.5 takes custom pixels: both edges a multiple of 16, a maximum edge of 3840, an aspect ratio of at most 3:1, and 655,360 to 8,294,400 pixels. A 21:9 frame fits; an 8:1 frame does not.

A worked strip order

Say you need 14 site-section dividers as 8:1 strips. On Nano Banana 2.1 at 1K that is 14 x $0.10 = $1.40; at 0.5K it is 14 x $0.075 = $1.05. Pro cannot produce the ratio natively, so the comparison is only meaningful if you accept a crop.

For a crop route, a 21:9 image on Pro at 2K costs $0.1875, and cutting a strip from it is your own step. Fourteen of those would be $2.62. The same dividers at native 8:1 on 2.1 cost $1.40 at 1K, so the native ratio is also cheaper.

Choosing

Pick the ratio first and the model second, because the list decides the model. If the ratio is in every list, pick by price and by how the text renders in your test. Ratio never changed the row on the banana models: tier does.

Read GET /v1/images/models/{model_id}/endpoints for the definitive descriptors before a batch. The generator cannot silently drop an unsupported ratio; it returns 400 unsupported_parameter.

Pitfalls with extreme ratios

Text in an 8:1 strip has very little height to work with. A single line of large type reads well; two lines do not. Generate the strip empty and set the words in your own layout step if the copy matters, since you then control the font and the kerning.

Do not send both a ratio and explicit pixels. The docs state that the ratio is ignored when image_size is set on models that take custom pixels, so send one of them. For the banana models, send aspect_ratio and the resolution tier, and leave pixel dimensions out.

If you need 21:9 on all three models for a comparison test, the three single images cost $0.15 + $0.1875 + $0.055625 = $0.39 at 2.1 2K, Pro 2K and GPT medium 2K. That is the price of a fair look at which one renders your subject best.