Someone asked me why a 2K job cost 112 credits when the rate card says ten credits a second and the clip was ten seconds.
The arithmetic looked wrong: 10 × 10 = 100, not 112. He was right to be suspicious, and the twelve missing credits were four reference images past the fifth, at three credits each.
That is the shape of every surprise in video billing. The per-second rate is real, and it is also only the first term.
The four terms
| Term | When it applies | What it adds |
|---|---|---|
| Tier rate × output seconds | every job | 1–10 credits per second depending on tier |
| Reference images past the fifth | 2K image-to-video | 3 credits per image |
| Video input seconds | video-to-video | billed at that path's rate, alongside output |
| Audio | reference clips and output track | nothing |
Term one: the rate everybody quotes
| Tier | Credits / output second | Ten-second job |
|---|---|---|
| 480P | 1 | 10 |
| 768P | 2 | 20 |
| 960P | 3 | 30 |
| 1080P | 3 | 30 |
| 1440P | 4 | 40 |
| 2K | 10 | 100 |
Term two: references, but only the sixth onward
The first five reference images are free on every tier. On the 2K image-to-video lane, each image beyond the fifth adds three credits.
Nine images on a ten-second 2K job: 100 + (4 × 3) = 112. A 12% increase that no per-second table shows, and the reason a "cheap" 2K experiment can quietly cost more than the plan said.
If you are budgeting, this is the term to design around: five well-chosen references cost nothing; nine mediocre ones cost twelve credits and dilute each other.
Term three: input seconds on a restyle
A video-to-video job bills the footage you feed it as well as the footage it produces — and that path's per-second rate is not the generation rate for the same tier.
Six seconds of source plus five seconds of output is eleven billed seconds, not five. Short sources are the cheap way to use the path; a fifteen-second source clip is not a small job.
Term four: audio, which is free, and failure, which is free
Reference audio (up to three clips) adds nothing to the bill, and every output already carries a native audio track.
Failed, cancelled and expired jobs release their reservation in full. You pay for output you can download, not for attempts — which changes the arithmetic of retrying: a retry is not a second purchase unless a file appears.
The arithmetic on a real plan
Twenty 5-second drafts at 480P (100 credits), then three 10-second finals at 1080P (90 credits): 190 credits, about $1.90, for twenty directions and three deliverables. The same plan rendered entirely at 1080P would be 300 credits before a single take was rejected.
Where this arithmetic does not apply
Free plans. Credits come from the weekly grant, so the cost is opportunity cost: 18 credits in week one is 18 seconds at 480P, or 9 at 768P.
Subscription pricing. The tier rates are identical on every plan, but the price of a credit varies with plan and cycle, so the dollar column is list rate rather than your rate.
2K with a large reference set. Add the per-image term before comparing 2K to 1440P, which shares the same 2560×1440 output for four credits a second.
FAQ
Do failed generations cost credits? No. failed, cancelled and expired jobs release the reservation.
Is audio input billed? No — up to three clips, plus the output track, cost nothing extra.
Why did a 2K job cost more than rate × seconds? Reference images past the fifth, at three credits each.
Does my plan change the per-second rate? No; it changes the price you pay per credit.
The live rate card is at Callirra's MiniMax H3 page. If your invoice shows a term this post does not cover, send the job configuration — that is a documentation bug I would rather fix than explain.
Top comments (3)
Good breakdown, and the generalisation is worth stating: the quoted rate is rarely the term that surprises you. Across the video and image APIs I have measured, the billed unit is per output second, per clip or per call depending on the tier, and reference and audio terms sit outside it. The practical fix is to price a job as a vector rather than a number - unit, units, tier multiplier, extras - so the 12% from four reference images is visible before you run it, and two models quoted in different units can still be compared fairly. The tier table is the other trap: 2K at 10 credits per second is ten times the 480P rate for a larger frame, and budgeting on the per-second number hides exactly that.
The 'per-second rate is only the first term' framing is exactly the surprise: 2K 10s at 100 credits plus four extra references is 112, and no rate card shows that.
The sixth-reference rule is the one to design around — five well-chosen refs cost nothing, nine mediocre ones cost 12 credits and dilute each other.