Nano Banana 2.1: 100 drafts at 0.5K plus 10 finals at 4K is $9.50

Nano Banana 2.1 prices by resolution tier on Sume, so a draft-then-final run is cheaper than shooting everything at the top tier. One hundred drafts at 0.5K cost $7.50 and ten finals at 4K cost $2.00, a total of $9.50. One hundred images straight at 4K would cost $20.00.

The tier ladder and the arithmetic are below, along with the one place the plan can go wrong: a draft at a small tier is a different sample from the final, not a preview of it.

The four tiers on Sume

The Image API docs accept resolution values 512, 1K, 2K and 4K (with 0.5K as an alias for 512) on models that list them. For Nano Banana 2.1 the catalog list rate is the 1K default, and admission scales the price for the other tiers. The billed amounts below are Sume's list x 1.25.

Working the plan

Drafts: 100 x $0.075 = $7.50. Finals: 10 x $0.20 = $2.00. Total: $9.50. Against 100 images at 4K ($20.00) you keep $10.50, which is 52.5%.

A 1K middle step changes the numbers only a little. Adding 30 images at 1K ($3.00) as a second filter before the 10 finals brings the run to $12.50. Whether that helps depends on how many of the 100 drafts you can reject by eye at the smallest size.

Where this can go wrong

A 0.5K image shows composition and colour but not fine detail such as small text or faces, so you may reject a concept that would have looked fine at full size, or accept one whose detail fails later. If the thing you care about is lettering or skin, draft at 1K instead.

Also remember that the Sume docs say seed is rejected with 400 unsupported_parameter. The 4K final will not reproduce the 0.5K draft; it takes the prompt again, or the draft as an input_references item if you want it to stay close.

Adding a middle step

The same pattern works with two filters. Generate all 100 at 0.5K, keep the 30 best, re-render those at 1K, and finish the 10 best at 4K. The extra step is a second look at a size where small flaws are visible, and it costs 30 images at the 1K price. If a quarter of the 1K candidates fail on detail, you have avoided spending the 4K price on them.

Keep a log with one row per generation: prompt id, tier, billed cost and verdict. After one or two runs you will know your own reject rate at each tier, and that number, not a guess, tells you whether the middle step pays for itself.

Before you run it

• Request resolution: "0.5K" for the drafts and confirm usage.cost reads $0.075.
• 4K requests can exceed the 30 second synchronous wait. Branch on the status code: 200 has the images, 202 has a job to read from the result endpoint.
• Set a ceiling in your own code. With 10 finals at $0.20, a loop that retries blindly could double the $2.00 final bill.