MiniMax H3 pricing: specs, costs and production fit

MiniMax H3 is a general-purpose multimodal video model built around a single context containing text, images, video, and audio. The practical headline is not simply that it generates 2K video. H3 can use one image for identity, another for a product, a video for motion, and an audio clip for voice or rhythm, then interpret those materials together in one generation request.

That flexibility also makes the pricing less obvious than a single per-second number suggests. MiniMax's official pay-as-you-go page lists H3 at $0.13 per second for 2K output and $0.08 per second for 768P output. Reference video is billed by its own duration as well as the generated output, while audio references and the first five image references do not add a material fee.

This guide separates the documented facts from the launch claims, works through the official price calculation, and shows where H3 fits in a production decision. It does not treat selected launch demos as proof of repeatability, and it keeps MiniMax's official API pricing separate from the live OmniArt credit schedule.

MiniMax H3 at a glance

MiniMax calls the API model MiniMax-H3. You may also see Hailuo 3.0 or Hailuo 03 on partner surfaces, but those labels should not be mixed with the older MiniMax-Hailuo-2.3 and MiniMax-Hailuo-02 models.

The full input limits, supported formats, file sizes, and request examples are documented in MiniMax's official H3 video-generation guide. Limits can change as the rollout expands, so production teams should check the live API documentation before locking a delivery plan.

Official MiniMax H3 pricing

MiniMax prices H3 output by the generated second.

For a generation without billable reference inputs, the base calculation is:

output cost = output duration × resolution rate

At the public 2K rate, that produces the following estimates.

These figures come from MiniMax's official pay-as-you-go pricing page, not a reseller or inference partner. They are list-price estimates before taxes, account-specific agreements, failed requests, or any future pricing change.

How reference inputs change the price

H3's reference billing rules matter when a brief uses motion transfer or a large asset pack.

For example, a 10-second 2K output directed by a 6-second reference video has a list-price estimate of:

output:    10 seconds × $0.13 = $1.30
reference:  6 seconds × $0.13 = $0.78
total:                           $2.08

The reference-video charge rewards careful asset preparation. Trim a motion reference to the exact section that carries the performance, camera movement, or editing rhythm you need. A long reference does not become more useful simply because the API accepts it.

If the same request uses seven images, the first five are free and the final two add $0.08. Audio references do not add a material charge, although they still need to meet the duration, format, and mixed-input rules in the API documentation.

What the multimodal context changes

Earlier AI video workflows often split a brief into separate operations: generate a silent clip, transfer motion, add dialogue, create ambience, then repair identity or product drift. H3 instead accepts those directions as parts of one context.

The useful production pattern is to assign one job to each input:

• A reference image protects identity or product appearance.
• A reference video supplies motion, camera behavior, or edit rhythm.
• A reference audio clip supplies voice, sound character, or timing.
• The text prompt describes the intended result and states what must remain unchanged.

That does not guarantee that every instruction will survive a crowded brief. It gives the model more explicit evidence to work with. Conflicting references can still produce a polished result that follows the wrong source, so a disciplined reference pack matters more than the maximum file count.

For practical reference-role patterns and reusable prompt structures, see the MiniMax H3 prompt guide.

API access is live; downloadable weights are not

MiniMax's official launch post says the company plans to release H3 model weights “in the coming days,” subject to applicable laws and regulations. At publication time on July 31, the API is live, but the weights are not yet available to download from MiniMax's public repositories and a final H3 license has not been published.

ModelScope's official MiniMax publisher page shows a scheduled release time of August 3 at 00:00 Beijing time. Treat that as a release target, not proof that the weights, code, license, and local deployment requirements are already available. Until the files land, “open-weight release announced” is accurate; “H3 is already open-source and runs on consumer hardware” is not.

For primary-source status, check the MiniMax H3 launch post and the MiniMax H3 ModelScope page. A useful local-deployment guide will need the actual license, weight size, memory requirements, inference code, and parity tests against the hosted API rather than launch-day estimates.

Where H3 looks most relevant

H3's documented input system points to several briefs where it deserves evaluation.

Reference-led product videos

A product team can combine clean product images, a motion reference, a visual-style reference, and audio direction. The important test is not whether the first frame looks premium. It is whether the label, silhouette, material, and product-to-hand interaction remain usable through the full clip.

Performance and motion transfer

Reference video can carry body movement, camera movement, or timing into a new subject and setting. This makes H3 relevant to fashion, dance, short-form narrative, and previsualization, but it also creates a billable input. Teams should compare the creative gain against the cost of the reference duration and the number of reruns required.

Native audiovisual drafts

H3 generates video with native stereo sound and accepts audio references. That can remove a separate temporary voice, ambience, or music pass during ideation. A client-ready mix still needs review for intelligibility, lip timing, spatial fit, rights, and consistency across edits. The H3 native-audio guide provides a structured way to test those layers without mistaking “audio present” for “audio finished.”

Graphic and brand work

MiniMax presents accurate text and brand rendering as a launch strength. Treat that as a testable claim. Use exact packaging, logo, and title-card references, retain every generation, and compare complete clips rather than selected stills. For client work, confirm trademark, likeness, voice, music, and source-footage rights separately from model capability.

What still needs independent testing

The launch material shows broad creative range, but a selected demo reel cannot answer production questions such as:

• How many runs are required before one is usable?
• Does product geometry survive a moving hand and camera at the same time?
• Does dialogue remain intelligible under ambience and music?
• Which instruction wins when image, video, audio, and text references disagree?
• Does a 15-second result hold identity and environment as reliably as a short clip?
• How do API latency and queue behavior change under real workloads?

A fair model comparison should hold the source assets, prompt, duration, aspect ratio, and review criteria constant. It should also keep weak outputs rather than publishing only the best generation. Until that evidence exists, describe H3 through documented capabilities and observed outputs, not universal rankings.

H3 cost per usable clip

List price is only the first layer of production cost. The more useful planning formula is:

cost per usable clip =
  (output cost + reference-input cost) × average attempts
  + external voice, music, cleanup, and finishing cost

A 10-second 2K clip with no billable reference inputs costs $1.30 at list price. If a brief averages four attempts before one clip is usable, generation spend is closer to $5.20. If each attempt also uses a 6-second reference video, the same four-attempt cycle becomes $8.32 before post-production.

That is why a cheaper model can be more expensive in practice if it needs many reruns, and why a higher-resolution generation can still be economical if it avoids upscaling or a separate temporary sound pass. Track accepted clips, rejected clips, causes of failure, and post-production time instead of comparing headline prices alone.

How H3 fits an OmniArt workflow

MiniMax H3 is available through PixVerse on OmniArt for Creator plans and above, with Video, Transition, and Reference modes, 5–15-second output, and 768p or 1440p (2K) resolution.

H3 costs 8 credits per output second at 768p or 12 credits per second at 2K. In Reference mode, reference video uses the same selected per-second rate; the first five reference images are included, each later image costs 4 credits, and reference audio adds no credits.

MiniMax's official dollar prices above and OmniArt credits are two separate billing systems; do not convert one directly into the other. Open MiniMax H3 on OmniArt to price the actual settings, then use the AI video model guide or photo-to-product-video workflow to prepare the brief.

MiniMax H3 FAQ

How much does MiniMax H3 cost?

MiniMax's official pay-as-you-go list price is $0.13 per generated second at 2K and $0.08 per generated second at 768P. Reference video is charged by its duration at the same resolution rate; audio and the first five images are free.

How much is a 15-second H3 video?

A 15-second 2K output without billable reference inputs is $1.95 at the official list price. Reference-video duration and images beyond the first five can increase the total.

Does a MiniMax Video Package include H3?

No. MiniMax's official Video Packages page currently states that H3 is not supported. H3 appears on the pay-as-you-go pricing page instead.

Is MiniMax H3 the same as Hailuo 3.0?

MiniMax-H3 is the official API model name. Some partner surfaces use Hailuo 3.0 or Hailuo 03 for the H3 route. Confirm the model ID and live documentation before assuming two differently named endpoints have identical settings or pricing.

Can H3 use image, video, and audio references together?

Yes. MiniMax documents mixed reference generation with up to nine images, three videos, and three audio clips, capped at 12 files total. Audio must be accompanied by a visual reference.

Is H3 available on OmniArt?

MiniMax H3 is available through PixVerse on OmniArt for Creator plans and above, with Video, Transition, and Reference modes, 5–15-second output, and 768p or 1440p (2K) resolution. MiniMax H3 on OmniArt

Is MiniMax H3 already open-source?

No downloadable H3 weights or final H3 license were public when this article was published on July 31, 2026. MiniMax announced an open-weight release for the coming days, while ModelScope shows a scheduled release time. Recheck the official repositories before making a local-deployment or commercial-use claim.