9-tile Instagram grid with 3 takes per tile: 27 images, from $0.67

A 9-tile grid with three takes per tile is 27 images: $2.70 on Nano Banana 2.1 at 1K or $0.67 on GPT Image 2.5 low on Sume. Arithmetic included.

4 min readSume
All posts

The short answer

A 3 by 3 Instagram grid with three takes per tile is 27 images. On Sume that costs $2.70 with Nano Banana 2.1 at 1K (27 x $0.10), $0.67 with GPT Image 2.5 low (27 x $0.02475 = $0.66825), and $1.50 with GPT Image 2.5 medium at 2K.

The three takes are what make the grid look planned: you pick one of three for each tile, so you pay for three and keep one.

The arithmetic

Nine tiles at three takes is the minimum for choosing. If you want a continuous picture across the grid, generate one wide image and crop it, which is one call at one price instead of nine.

The table prices the full 27. The 4:5 ratio is Instagram portrait, and the docs say it is not 4:3.

27 images for a 9-tile grid, as of 2026-10-08 (totals rounded to the cent; the Google row is a reference and not sold through Sume)
OptionCount x unit priceTotal
Nano Banana 2.1 at 1K27 x $0.10$2.70
Nano Banana 2.1 at 0.5K27 x $0.075$2.03
GPT Image 2.5 medium (2K)27 x $0.055625$1.50
GPT Image 2.5 low (1K)27 x $0.02475$0.67
Google direct, 1K standard (not on Sume)27 x $0.0336$0.91

How to run it

Request three takes in one call by setting n to 3, which keeps the prompt identical across takes. Sume charges cost_usd times n, so the bill is the same as three single calls.

For exact 1080 by 1350 output, note the docs: Nano Banana sends the 4:5 ratio and an exact size comes from a documented post-step, while GPT Image 2.5 needs edges that are multiples of 16, which 1080 is not.

Retries and cost per tile

Some images will be rejected by eye, and rejected images are billed like any other, so budget a margin. The bullets add 20 percent to each Sume total in the table above (the totals multiplied by 1.2) and divide by the number of tiles to give a per-tile cost.

Treat the per-tile figure as the number to compare across models, because it does not change when the count changes. Re-read the public catalog before a large run, since prices can change and the table is a snapshot from 2026-10-08.

  • Nano Banana 2.1 at 1K: $2.70 becomes $3.24 with 20 percent extra images, which is $0.36 per tile (a 9-tile job).
  • Nano Banana 2.1 at 0.5K: $2.03 becomes $2.44 with 20 percent extra images, which is $0.27 per tile (a 9-tile job).
  • GPT Image 2.5 medium (2K): $1.50 becomes $1.80 with 20 percent extra images, which is $0.20 per tile (a 9-tile job).
  • GPT Image 2.5 low (1K): $0.67 becomes $0.80 with 20 percent extra images, which is $0.09 per tile (a 9-tile job).

What the docs say to check

The endpoint pricing lines are what Sume charges your wallet and already include the Sume margin, so a call costs cost_usd times n. Sume bills image generation all or nothing: a generation that fails or is cancelled is not charged.

Read GET /v1/images/models before you pin a tier or ratio. A model accepts only the values its descriptors list, and the descriptors are the definitive source for what a call may send.

Put the three takes of each tile side by side before you choose, since a grid is judged by how tiles sit next to each other.

Every price in this post is a snapshot dated 2026-10-08. The Sume figures come from the public catalog and the Image API docs, the Google figures come from its Gemini API pricing page, and no figure here is a benchmark of quality or speed. Prices and model lists change, so confirm them on the day you spend money.

Sources

Related posts

More in Use cases

All Use cases posts

Written by Sume