Nano Banana 2.1 or Pro? The billed gap at 1K, 2K and 4K on Sume

On Sume, Nano Banana 2.1 bills about $0.10 per 1K image and Nano Banana Pro about $0.19, so Pro costs roughly 1.9 times as much at 1K. At 2K the gap narrows to 1.25 times ($0.1875 against $0.15), and at 4K it is about 1.9 times again ($0.375 against $0.20). Both numbers are Sume's catalog list times 1.25, taken from the repository pricing tables on 2026-10-08, so confirm them in the pricing field of the endpoints reply before you budget.

Google made Nano Banana 2.1 generally available on October 6, 2026, which puts the cheaper row on the default path for most image work. Pro stays in the catalog. The question is which jobs justify the extra cents.

Side by side on Sume

Where Pro still makes sense

Pro has the same ratio set minus the four extreme banner values, so you do not choose it for shape. The honest reasons are quality on a hard prompt and consistency you have already tested. Run your own prompt on both rows: one request each costs about 29 cents at 1K, and you see the difference before you commit a campaign to either.

A quick decision rule

• Banners at 4:1, 1:4, 8:1 or 1:8: only 2.1 lists them.
• Drafts at 0.5K: 2.1 bills $0.075 against $0.1875 on Pro, so drafts belong on 2.1.
• A final hero image where a test shows Pro is better: pay the extra $0.0875 at 1K.
• Everything else: start on 2.1 and keep Pro as the fallback.

Cost of testing both

A fair test uses the same prompt, the same ratio and at least five prompts, because one image proves little. Five prompts on both rows at 1K cost five times $0.10 plus five times $0.1875, which is about $1.44. If the output is close, you have saved real money on every later batch; if Pro wins on your product shots, you now know the premium buys something.

Use aspect_ratio: "auto" when you edit from a reference photo so both models keep the source shape, and keep n at 1 while you compare.

Reading the endpoints reply

Both rows publish a single sume endpoint, so the model-level and endpoint-level parameter lists are identical. The pricing array holds the billable line, for example output_image per image, and the cost already includes Sume's margin. If the table above ever disagrees with that field, the field wins.

Because 2.1 and Pro share the same resolution enum, you can switch rows by changing only the model string and re-reading the price. That makes a one-line fallback easy: try 2.1, and send the same body to Pro only for prompts a reviewer rejected.

What Sume does not do

Sume does not rank the two models for you and publishes no quality benchmark. Sume also caps input_references at 10 on both rows, while Fal's page for the 2.1 edit endpoint lists up to 14, so a 14-image job needs trimming.