Cheapest 4K image on Sume: GPT Image 2.5 $0.014 vs Nano Banana

The cheapest 4K image on Sume is GPT Image 2.5 at 3840x2160 and low quality, $0.0140, followed by medium at $0.0325. Nano Banana 2's 4K tier is $0.2000 and Nano Banana Pro's is $0.3750, which is 6.2 times the medium price and 14.3 times the low price. At high, GPT Image 2.5 lands at $0.1251, still below both Nano Banana rows.

Those are different products, so the price is the start of the decision and not the end of it. Both families can reach 4K, but they do it through different parameters, and the cheaper row also needs the more careful request, as the sections below explain. Every figure here is Sume's quote and includes the 1.25 factor over the Fal list price.

4K prices on Sume

GPT Image 2.5 is priced by output tokens, so every tier has its own 3840x2160 price. Nano Banana rows are flat per resolution tier. All figures include Sume's 1.25 factor.

What 4K means in each row

On GPT Image 2.5, 3840x2160 is a custom size: both edges are multiples of 16, the longest edge is exactly the 3840 maximum, and the pixel count is 8,294,400, the top of the allowed range. Sume's Image API page lists those rules. Nothing larger is accepted.

On Nano Banana the 4K value is a resolution tier, set with resolution, and the aspect ratio is separate. Sume's catalog does not publish pixel dimensions for the tier, so check the file you get back instead of assuming. Nano Banana 2 lists 15 aspect ratios in Sume's catalog and Nano Banana Pro lists 11.

The request

A 4K GPT Image 2.5 call at medium quality. Expect a slow render: Sume waits 30 seconds and returns 202 with a job envelope for anything longer, and large sizes at high quality are the likeliest to do that.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Aerial view of terraced rice fields at dawn",
    "quality": "medium",
    "image_size": "3840x2160",
    "mode": "async"
  }'

When the cheap row is the wrong row

Pick on features after price, because the lowest quote is only useful if the model can do the job. Only the GPT Image 2.5 rows document background: transparent. Nano Banana takes a resolution tier and an aspect ratio rather than pixels, and both Nano Banana 2 and Nano Banana Pro accept up to four output images per call.

If you need 4K only for print or a large display, the low tier at $0.0140 lets you test a layout at full size for a fraction of a cent before the final render. If you need Nano Banana's ratios or look, 4K costs $0.2000 on Sume and there is no cheaper 4K tier.

Nano Banana Pro is the one row where 4K doubles the price: its 1K and 2K tiers both quote $0.1875 on Sume and 4K quotes $0.3750. If you were considering Pro at 2K, 4K is the only step that changes the bill.

Checking the file you get back

Whatever the row, read the dimensions of the returned image before you build a pipeline on it. The 4K promise is a billing tier on one side and an exact pixel size on the other, and a mismatch between them is easy to miss when the preview is a thumbnail.

For GPT Image 2.5 the returned size is the one you asked for, within the limits above. For Nano Banana, pass an aspect_ratio explicitly instead of relying on auto, because an edit that omits the field does not behave like auto. Run the 4K size only on images you have already approved at a smaller size, since the cost of a rejected 4K render is higher than that of a rejected draft. Keep one test call per row in your own account and compare usage.cost with the table before you commit a batch. Failed generations are not billed, so a test that errors out costs nothing, but a successful one does.