Nine reference images: H3 adds $0.40 on Sume, H3 Max adds none

On Sume, minimax-h3 reference-to-video adds $0.08 list per reference image after the first five, which is $0.10 after the 1.25 multiple, so a nine-image brief adds 4 x $0.10 = $0.40. minimax-h3-max is billed at its per-second list, so the same nine references add nothing to the reservation.

That makes H3 Max the cheaper row for large reference sets once the extra images outweigh its higher per-second price. At 10 seconds and 768p, the crossover is at eight images.

What is documented

The Video Router catalog note for minimax-h3 reads: r2v +$0.08 per image after 5, with fal list $0.05 and $0.06 per second for native 480p and 768p. The pricing source for H3 Max says its r2v sells the same per-second list, and that fal's reference-token overage is not reserved. Both rows accept reference_*_urls in the Video Router.

On the vendor side, the MiniMax announcement (read 2026-10-05) describes reference as a general capability across image, video and audio without giving counts on that page. A third-party summary (read 2026-10-05) lists up to 9 reference images, 3 video clips and 3 audio clips for H3; treat that as secondary and confirm with the catalog.

Price table with nine images

The table assumes 768p. The H3 total is the per-second price times the length plus $0.10 for each image after the fifth.

How to use the table

Count the references you really need. If the extras only restate the same product from the same angle, the cheapest fix is to drop them. If each image carries a different person, product or location, the cost is the price of a more controlled output.

For short clips the crossover is early because H3's discount is small in absolute terms. A 5-second clip is $0.375 on H3 and $0.50 on H3 Max, a gap of $0.125, so just two extra images erase it. At 15 seconds the gap is $0.375, which takes four extra images to erase.

• Up to five images: H3 is cheaper at every length.
• Six to seven images: H3 is still cheaper at 10 s and 15 s.
• Eight or more: H3 Max is cheaper at 10 s; nine at 15 s.

Caveat on rounding

Sume rounds the billable total up to the cent, so the table values can differ by one cent from a hand calculation. The submit response carries usage.billable_amount_usd_micros; that is the number that counts.

Choosing references that earn their price

A reference image changes the output only when it shows something the prompt cannot say: a specific face, a pack design, a room. Three images of the same pack from three angles can carry more than six near-duplicates. On minimax-h3 each image after the fifth costs $0.10, so six well-chosen images are cheaper than nine redundant ones by $0.30.

The Sume catalog also reports which reference types a row accepts. For H3 and H3 Max, the docs list image, video and audio references, and supported_input_references on the row tells you the current set. Read it before you build a brief around a type.

Test with the cheap row first

If you are unsure whether the extra references help, run one 480p take on each id. At 480p both rows bill $0.0625 per second after the multiple, so a 10-second take is $0.63 plus any H3 surcharge. Compare outputs and add the images that moved the result.