63-slide deck, one 16:9 image per slide: $1.56 to $9.45 on Sume

One 16:9 image for each of 63 slides costs $1.56 on GPT Image 2.5 at low quality, $2.76 on Seedream 5 Lite, $6.30 on Nano Banana 2.1 at 1K and $9.45 at 2K, using Sume's per-image prices as of 2026-10-09. Every row in the table lists 16:9, so the choice is about price and look, not ratio.

A 63-slide count is a real planning number: a one-hour talk with a slide a minute, plus a title, an agenda and a closing slide. Few decks need an image on every slide, so treat these totals as the upper bound.

Totals by row

The third column is the 63 images multiplied by the per-image price. The fourth column shows what the balance holds if you submit 63 separate single-image calls, since Sume rounds each hold up to whole cents.

Why the hold differs from the bill

Sume reserves the estimated amount at submit time, rounded up to whole cents, and bills the actual completed generation. A GPT Image 2.5 low image lists at $0.02475 and the hold is 3 cents. Over 63 calls that is $1.89 held against $1.56 billed. The extra is released as each job completes, but a small wallet needs the headroom at the moment you submit.

Batching with n changes the rounding. Seven calls of n: 9 on GPT Image 2.5 low hold 23 cents each, since nine images cost $0.22275, so the total hold is $1.61. The Image API docs explain n and the per-model ceilings, and some models cap n below ten.

A practical split

Most decks need a style, not 63 finals. Draft the whole deck on GPT Image 2.5 low for $1.56, pick the ten slides that carry the talk, and regenerate those ten on Nano Banana 2.1 at 2K for $1.50. The total is $3.06, against $9.45 for 63 images at 2K.

Keep the prompt prefix identical across slides so the set reads as one deck, and put the slide number in the job metadata so you can match each result to a slide. Sume stores metadata on the job and does not send it to the provider.

Check the deck for text. Slides with words on the image are the usual failure, so keep the words in the slide and ask the model for a picture without lettering. If a slide must carry text in the image, test it on one slide before running all 63, and re-run only the slides that fail. Because Sume bills only completed generations, a failed call costs nothing, but a successful image with the wrong text does.