H3 Max lip sync on a 6.2-second line: billed as 7 s, $0.44 to $1.40

A 6.2-second voice line sent to MiniMax H3 Max Lip Sync on Sume is billed as 7 seconds, because the catalog prices the route per ceil of the audio second. That is 7 x $0.0625 = $0.4375 at 480p, 7 x $0.10 = $0.70 at 768p (the default) and 7 x $0.20 = $1.40 at 1080p, as of 2026-10-09.

A 6.2-second line is a useful test case because it is just above the 5-second minimum the route accepts. Anything under 5 seconds is refused rather than padded.

The three resolutions

The route takes a still image (or a ready avatar) plus Sume-hosted audio, and returns a clip whose mouth follows the audio. Output length follows the audio. The resolution field accepts 480p, 768p or 1080p, with 768p as the default and no 2K option.

Window rules that decide whether the request is accepted

The OpenAPI schema for this route sets duration_seconds between 5 and 14.8, and says the provider rejects shorter audio and silently clips longer audio. Sume therefore refuses requests outside the window instead of clamping them, so you never pay for a clip that was silently cut. The still's aspect ratio must be between 0.4 and 2.5. The audio URL must be on the Sume media host and under 10 MB; URLs on other hosts are rejected.

Declare duration_seconds as the measured length of the audio. The reservation is ceil of that value times the resolution rate times the standard margin, so declaring 6.2 reserves 7 seconds of rate. Declaring 7 gives the same reservation, and it is the safer habit because it matches the billed unit.

• Audio length: 5 to 14.8 seconds.
• Still aspect ratio: 0.4 to 2.5.
• Audio: Sume-hosted, under 10 MB.
• speed_tier is accepted and ignored.

Draft cheap, finish sharp

Because the three resolutions share one body, a cheap pass costs little. Ten takes of this line at 480p come to $4.375. Ten at 1080p would be $14.00. One reasonable habit is to check timing and expression at 480p, then rerun only the keeper at 1080p. The extra for the final render is $1.40 - $0.4375 = $0.9625 for this line.

If the line runs longer than 14.8 seconds, split the script at a sentence boundary and send two jobs, or use the Fabric route, which accepts up to 300 seconds at a higher per-second rate. For the exact boundary behaviour see the 14.8-second case.

A quick budget table for a campaign

Say you need 40 lines of about 6 seconds each for a localization test. At 480p that is 40 x $0.4375 = $17.50 if each bills as 7 seconds. At 768p it is 40 x $0.70 = $28.00. At 1080p it is $56.00. Run the whole set at 480p, review them, then spend $56.00 only on the subset you keep. If 12 of the 40 survive review, the 1080p pass costs 12 x $1.40 = $16.80 on top of the $17.50 draft pass, $34.30 in total, versus $56.00 for rendering everything at 1080p.

Why the minimum is five seconds

The route refuses audio under 5 seconds. A 4-second line is not padded to five and billed, it is rejected, and the schema says why: the provider rejects shorter audio. For short lines, use VEED Fabric 1.0, which accepts 1 to 300 seconds at $0.10 (480p) or $0.1875 (720p) per second. A 4-second line there is 4 x 0.1875 = $0.75 at 720p.