MiniMax H3 Pricing and API Cost Guide: Output, References, and Retries

MiniMax H3 pricing has two valid answers. MiniMax publishes direct pay-as-you-go rates for 768P and 2K output. ClipDance currently offers H3 through a third-party provider route with fixed 2K output, a different per-second rate, and ClipDance credit conversion.

Mixing those routes produces misleading estimates. A direct 2K six-second output is $0.78 before references. The same duration on the current ClipDance route has $1.095 of provider usage and rounds to 55 ClipDance credits, equivalent to $1.10 at the configured $0.02-per-credit value. Neither number describes the other product's bill.

All prices in this guide were checked on August 13, 2026. Model access, provider rates, and credit packages can change, so calculate again from the checkout or console you will actually use.

MiniMax H3 pricing at a glance

The current MiniMax H3 generator converts provider cost into whole credits using this rule:

credits = ceil(provider cost / $0.02)

Credits round up at the request level. That is why four seconds of bare output costs $0.73 at the configured provider rate but becomes 37 credits, a $0.74 credit value. This is a UI and billing conversion for ClipDance, not MiniMax's official credit system and not a public ClipDance API price.

Output-only costs for 4, 6, 10, and 15 seconds

For text-to-video, image-to-video, or reference-to-video with no paid inputs, duration supplies most of the bill.

The ClipDance calculator uses the exact $0.1825 provider rate even though a public provider card may display the rounded figure $0.183. Calculations that multiply the rounded card price can therefore differ by a cent from the final whole-credit result.

These are generation charges, not complete project costs. They exclude failed creative attempts that complete successfully but are rejected by your team, human editing, storage, taxes, and any difference between the configured credit value and the credit package you purchased.

How image, video, and audio references change the price

Up to five images do not add a reference fee

Both routes include the first five reference images without a separate image charge. A sixth through ninth image costs $0.04 each through MiniMax direct or $0.055 each on the current ClipDance route. Using all nine therefore adds $0.16 direct or $0.22 through the current route.

That does not mean five images are always cheaper in practice than two. Conflicting identity, wardrobe, lighting, and style references can lower the acceptance rate. Add an image only when it supplies a fact the existing set cannot: another product angle, a missing costume detail, or a clearly different role.

Video references are billed as time

MiniMax direct bills reference-video duration at the output resolution's rate. Ten seconds of input video adds $0.80 to a 768P request or $1.30 to a 2K request.

On the current ClipDance route, reference-video seconds are added to the requested output duration. Input duration is rounded up to a whole second, and total billable time is capped at 30 seconds. A six-second output with 9.2 seconds of reference video is therefore billed as 16 seconds, not 15.2.

Reference videos may be worth the charge when exact body motion, performance timing, or camera movement is hard to specify in words. They are expensive decoration when the prompt already describes a simple pan or walking pace.

Audio input is free, but it cannot stand alone here

Neither route adds a media fee for an audio reference. On the current ClipDance reference workflow, however, audio cannot be submitted as the only reference. It needs at least one image or video. H3 generates native stereo sound on the current route; there is no cheaper silent-output toggle.

For practical reference selection and role-assignment examples, use the MiniMax H3 prompt and reference workflow.

MiniMax direct 768P versus 2K

MiniMax direct lets developers choose 768P at $0.08 per output second or 2K at $0.13. A 15-second draft is $1.20 at 768P and $1.95 at 2K. The current ClipDance route is fixed at 2K, so that resolution choice is unavailable in its form.

MiniMax also lists a 768P-to-2K regeneration rate of $0.05 per output second. The original input materials are billed again during regeneration: audio remains free, each image after the fifth costs $0.025, and reference video costs $0.05 per second. For output and video input, a 768P run followed by regeneration reaches the same combined per-second rate as starting at 2K. Extra images make the two-stage path slightly more expensive because their original and regeneration charges stack.

Use 768P first when it can answer a composition or motion question without a guaranteed upgrade. If every accepted draft must become 2K, generating directly at 2K is simpler and avoids a second task and another opportunity for cost confusion.

API mode is selected by the request shape

H3 supports text-to-video, frame-guided image-to-video, and mixed-reference video. The provider route behind the current ClipDance form uses one model id and decides the mode from the media fields. The fragments below document that current integration logic. They are not a public ClipDance v1 API contract, and they should not be copied into MiniMax's official API without checking MiniMax's current schema.

Text-to-video sends a prompt, duration, and required explicit aspect ratio:

{
  "model": "minimax-h3",
  "prompt": "A single continuous product shot...",
  "duration": 6,
  "aspect_ratio": "16:9"
}

Frame-guided image-to-video sends a first frame, an optional last frame, and no aspect ratio. Orientation comes from the source image:

{
  "model": "minimax-h3",
  "prompt": "Move naturally from the opening frame to the closing frame...",
  "duration": 6,
  "first_frame_url": "https://example.com/opening.jpg",
  "last_frame_url": "https://example.com/closing.jpg"
}

Reference-to-video sends at least one image or video. It can also send audio, but audio alone is rejected:

{
  "model": "minimax-h3",
  "prompt": "Use Image 1 for identity and Video 1 for camera movement...",
  "duration": 10,
  "reference_image_urls": ["https://example.com/identity.jpg"],
  "reference_video_urls": ["https://example.com/motion.mp4"],
  "reference_audio_urls": ["https://example.com/voice.mp3"]
}

The current integration accepts up to nine images, three videos totaling no more than 15 seconds, and three audio files. It requires a 1–7,000-character prompt and a whole output duration from 4 through 15 seconds. Text mode rejects adaptive as an aspect ratio; reference mode may omit a ratio for adaptive framing; frame mode must omit it.

Mode mistakes are cost-control problems even when rejected validation is not billed. They waste engineering and review time. Decide whether the source media is a boundary frame or a reusable identity/motion reference before assembling the payload.

Budget the accepted clip, not the first generation

The per-second row is a unit cost. Production cost depends on how many completed outputs you reject.

cost per accepted clip = total generation spend / accepted clips

Suppose one usable six-second shot requires four completed attempts:

Now add a ten-second motion reference to every attempt. The current route becomes 146 credits per run, or 584 credits ($11.68 value) for four completed attempts. That is why a motion clip should be introduced after a cheaper short test proves the subject, scene, and basic action.

Track accepted cost by shot type rather than averaging an entire campaign. Dialogue, readable packaging, multi-person action, and identity-heavy clips fail for different reasons. A model may be cheap for establishing shots and expensive for close product work even though its rate never changes.

A minimal production log needs only the route, mode, duration, paid input seconds, paid image count, credits or dollars charged, pass/fail result, and one rejection reason. Ten rows of honest results are more useful for budgeting than a selected comparison clip. If you are choosing between models rather than routes, the existing Seedance 2.5 vs MiniMax H3 comparison explains what a controlled evaluation can and cannot prove.

Which route makes sense?

Choose MiniMax direct when your team wants MiniMax's own API relationship, needs the 768P option, and can implement and operate the official contract. The published list price is lower, but your real decision should still include payment access, queue behavior, support, integration work, and retry rate.

Use the current ClipDance H3 form when browser access and a unified creative workflow matter more than the lowest raw rate. Its fixed 2K provider route and whole-credit conversion cost more than MiniMax's direct list price. ClipDance is an independent platform, not MiniMax, and the difference should be evaluated as a different access product—not described as MiniMax's price.

For either route, recheck the live price on the day you commit a batch. Save the rate, payload settings, and reference durations with the job record. A cost estimate without its date and access route is not reproducible.

References

1. [1] MiniMax API Docs. Pay as You Go Pricing. Accessed August 13, 2026. Official output, input-material, and regeneration list prices.
2. [2] MiniMax. MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities. Published July 31, 2026. Official capability and launch context.
3. [3] reAPI. MiniMax H3 model page and current provider pricing. Accessed August 13, 2026. Public provider display rounds its per-second rate; ClipDance's current calculator uses the configured exact rate and its own credit conversion.