2,500 Nano Banana 2.1 images at 1K: Google Batch $42 vs Sume $250

For 2,500 Nano Banana 2.1 images at 1K, Google's Batch tier costs $42.00, Google's Standard tier costs $84.00, and Sume costs $250.00. Google's rates come from its Gemini API pricing page and Sume's from its catalog row, both read on 2026-10-09.

The gap is real, so the useful question is what you are buying with the difference rather than whether Sume is cheaper (it is not, for this one model).

The arithmetic

Google lists gemini-nano-banana-2.1 at $0.0336 per 1K image on Standard and $0.0168 on Batch. Sume's Nano Banana 2.1 row at 1K is $0.10.

What Sume has and does not have

Sume's Image API docs describe a synchronous call that blocks for up to 30 seconds, then a 202 job envelope for slower work. They describe no discounted batch tier, so there is no Sume number to put next to Google's $0.0168. Batch is a Google feature, and half-price work in bulk is the one case where calling Google directly is the clear choice.

What Sume does document is the following.

• One request shape, POST /v1/images, across the models in the catalog, so changing a model is a changed model string.
• Results come back as Sume-hosted URLs, not inline base64.
• A generation that fails or is cancelled is not billed; a completed one is billed in full.
• Wallet billing in one place, including the other media models (video, music, speech) that the same account can call.

A decision rule

If the job is 2,500 identical-model images that can wait for a batch window, Google Batch at $42 wins on cost. If the job mixes models (a Nano Banana 2.1 draft, a GPT Image 2.5 edit, a background removal), or you need results in seconds through one integration, the per-image premium buys a single bill and a single client.

A middle path is to prototype on Sume, where you can switch models cheaply, and move one stable high-volume model to the vendor's batch API once the prompt is frozen. Keep the Sume path for everything else.

Before you pick on price alone

Direct access also means a separate account, a separate key, and a separate way of waiting for results, since a batch job is asynchronous by design. Sume's synchronous call returns in up to 30 seconds for most catalog models, which is a different product shape from a batch window.

Count the engineering time too. If wiring a second vendor takes a day and the saving is $208 (250 - 42), the second integration pays for itself only if you will repeat the run several times. For a one-off 2,500-image job, either path is fine; for a monthly job, run the numbers on 12 months: 12 x $208 = $2,496 saved by Batch.