H3 or H3 Max? Pick the Sume row by job, with per-clip prices

Choose minimax-h3 for the cheapest native 768p clip and when you may need its 2K/4K upscale rows; choose minimax-h3-max when you want a 1080p output or the faster 768p variant, at $0.10 per second instead of $0.075.

Both ids make 5 to 15 second clips with native stereo audio, take text, first/last frame, and reference inputs, and bill list x 1.25 on Sume. The difference shows up in resolution rows, in how reference images are charged, and in your per-second budget.

Side by side

The Video Router docs list both rows with the same duration range. The differences below come straight from the catalog notes (fal list pages read in the repo docs, read 2026-10-05).

Pick by job

A product teaser that will be cropped for a phone feed rarely needs 1080p on the first pass. A 10-second 768p take is $0.75 on H3 and $1.00 on H3 Max, so ten exploratory takes cost $7.50 versus $10.00. Run the explorations on H3, then re-render the winner on Max when you need the 1080p row.

For a hero shot with several references, remember the r2v rule. Sume's docs state that H3 charges an extra $0.08 list per reference image after the fifth, which is $0.10 per image after the 1.25 multiple. The Max row is described as billed at the per-second list. If your brief carries eight reference images, that difference matters.

• Cheap tests at 480p: both ids are $0.05 list, $0.0625 per second billed.
• Final 768p: H3 at $0.075 per second, Max at $0.10.
• Final 1080p: only Max, $0.20 per second, so 10 seconds is $2.00.
• Larger than 1080p: only the H3 upscale rows, and only if the live catalog still lists them.

What neither row does

Neither id goes past 15 seconds, so a 30-second ad is two clips plus a join. Neither has an audio toggle, so a silent take needs a different model. MiniMax's own H3 page (read 2026-10-05) advertises native stereo sound, which matches what the Sume rows say.

Finally, read capabilities from the catalog before you hard-code any row. The docs say the limits differ by model, and the endpoint is the contract.

A budget example

Suppose a campaign needs six finished 10-second clips and you test ten takes per clip at 768p. On H3 that is 60 takes x $0.75 = $45.00. On H3 Max it is 60 x $1.00 = $60.00. Refining the six winners to 1080p on Max adds 6 x $2.00 = $12.00 if you only pay for the finals, for a total of $72.00 when exploration ran on Max, or $57.00 when exploration ran on H3 and only the finals ran on Max.

The numbers are from the per-second rates in the table; the split assumes you accept a visual difference between the two ids during exploration. If prompt behavior on Max matters to the final, run your exploration there.

Polling and delivery are identical

Both ids go through the same Video Router submit and the same GET /v1/jobs/:id/status poll. The choice of id changes the price and the catalog row, not the lifecycle. The submit response carries usage.billable_amount_usd_micros, so log it per job and you can check the table above against your own ledger.