37 listing hero images at 4K: $7.40 on Nano Banana 2.1 via Sume

Thirty-seven 4K hero images cost $7.40 on Sume with Nano Banana 2.1 ($0.20 each), $8.23 with GPT Image 2.5 at high quality ($0.2225 each) and $13.88 with Nano Banana Pro ($0.375 each). Those are the catalog rows read on 2026-10-09; failed generations are not billed.

The count of 37 is a stand-in for a small shop that is replacing one hero picture per product page. Swap in your own count: the cost is the count times the per-image row, with no per-size surcharge beyond the tier you pick.

The three 4K rows on Sume

Sume bills image generation per image, and the price depends on the model and the resolution tier. For 4K the three closest choices are below.

What Google charges for the same 4K image

Google's own Gemini API page lists Nano Banana 2.1 (gemini-nano-banana-2.1) at $0.113 per 4K image on the Standard tier and $0.0567 on Batch (read 2026-10-09). For 37 images that is 37 x $0.113 = $4.18 Standard, or 37 x $0.0567 = $2.10 Batch.

So the direct Google rate is lower than Sume's $0.20 row. What Sume adds is the wallet and the job layer: one POST /v1/images, Sume-hosted result URLs instead of inline data, and an all-or-nothing bill where a failed generation is not charged. If you already run a Google account and only need one model, calling Google is cheaper. If you switch models between runs, one wallet and one request shape is the reason to use Sume.

Run it

Send one request per image, or use n where the model's descriptor allows it. GET /v1/images/models lists the allowed n and resolution values for each model, and Sume answers 400 unsupported_parameter for anything the model does not list. A request blocks for up to 30 seconds and returns a 202 job envelope if a 4K render is slower, so poll GET /v1/jobs/{id}/status and then fetch the result.

Budget check: multiply by the number of images you will keep, not the number you will try. If you expect to discard one in three, 37 keepers is about 56 attempts, or 56 x $0.20 = $11.20 on Nano Banana 2.1.

Checks before a 37-image run

Run one image first and read the response usage.cost. Sume returns the billed USD amount there, so you can confirm that the per-image row matches the table before you multiply it by 37. Token counts in that object are always 0 in v1, because Sume meters image models per image.

Keep the wallet above the total. Sume reserves an estimate when a job is submitted and captures it on completion, so 37 queued 4K jobs hold about $7.40 until they finish. A failed job is refunded before capture.

Prices on the catalog can change, and both vendors list their own rates on their own pages; recheck the Sume row with GET /v1/images/models on the day you run.