MiniMax H3 Max reference-to-video: what the reserve leaves out
Sume bills H3 Max reference-to-video by output seconds at list x 1.25; the provider's reference-token overage is not reserved. What to check before a big batch.

For minimax-h3-max reference-to-video, Sume's catalog says the job is billed as output seconds at list price times 1.25, and that the provider's reference-token overage is not reserved. In plain terms: the hold at submit is built from seconds and resolution, not from how heavy your reference files are.
That is a modeling note for a budget, not a reason to avoid references. It means you should treat the reserve as the main number and verify the final charge on a few real jobs before you run hundreds.
What is reserved and what is billed
Sume admits a video job in three steps: admit, reserve, capture. The reserve holds list times 1.25 for the request you submitted. The captured amount is what you read back as usage.cost on the poll response.
| H3 Max resolution | List per second | Billable per second (list x 1.25) |
|---|---|---|
| 480p | $0.05 | $0.0625 |
| 768p | $0.08 | $0.10 |
| 1080p | $0.16 | $0.20 |
A safe way to test
Run three jobs with the same prompt and different reference loads: one image, five images, nine images plus a short video. Read usage.cost on each completed job and compare it with the number the catalog quote predicts. If the three match, your budget model holds. If they drift, add a margin to the batch estimate instead of trusting the per-second math alone.
Pin the resolution while you test. Going from 768p to 1080p raises the rate on its own, which would hide a reference-related change.
Keep the balance above the reserve
A workspace with a balance under the reserve gets a 402 insufficient_credits at submit, before any compute runs. A failed job refunds its reserve. Keep a buffer for the part the reserve does not cover, and stop a batch on the first 402 instead of retrying through it.
What to log per job
A small ledger makes the gap visible without any extra tooling.
The point of the log is to turn an unknown into a measured number. After twenty jobs you will know whether reference weight moves your cost, and you can put a margin on the estimate based on data.
- The job id, resolution and duration you requested.
- The number of reference images, videos and audio files.
usage.costfrom the completed job.- The quote you computed before submitting, so you can see the difference per job.
Sources
Related posts
More in Pricing
- Monthly AI content budget at three sizes: starter, growth and scale
Three monthly volumes of images, voiceovers and 6-second clips priced on Sume: the starter plan is about $10, growth about $39 and scale about $146.
- Budget for 10 AI videos a week: 15-second Wan 3.0 clips, 3 resolutions
Ten 15-second Wan 3.0 clips a week cost $9.40 at 480p, $18.80 at 720p and $37.50 at 1080p on Sume. Per-clip, per-month and per-plan numbers.
- Omni Flash: Google's $0.10 a second vs Sume's $0.125 at 10, 30, 60 s
Google lists Gemini Omni Flash at about $0.10 a second at 720p. Sume bills $0.125. That is $0.25 more on 10 s, $0.75 on 30 s and $1.50 on 60 s. Read 2026-10-06.
- Omni Flash: $9 per 1M text tokens vs $17.50 per 1M video tokens
Google lists Gemini Omni Flash at $9.00 per 1M text output tokens and $17.50 per 1M video tokens. 2,000 text tokens cost 1.8 cents, about 0.18 s of video.
Written by Sume