What a MiniMax H3 clip really costs for legal presentations

As last verified on July 31, 2026, MiniMax H3 is priced at $0.13 per second for 2K output and $0.09 per second for 768P output. That makes the clean invoice math almost absurdly small: a 15-second H3 clip costs $1.95 at 2K or $1.35 at 768P before any input-material charges, retries, review time, preservation work, or courtroom explanation enters the file. [1]

For a legal presentation team, that is the wrong place to stop. H3 is not currently covered by MiniMax prepaid Video Packages; the package documentation says “MiniMax H3 is not supported yet,” which leaves H3 in pay-as-you-go billing rather than a predictable monthly pool. [2] The lower 768P tier also comes with access friction: public H3 writeups describe 768P generation as a closed beta requiring sales contact, so the $1.35 example is not the same thing as universally available self-serve production. [3][4]

The clip is cheap; the billing structure is not trivial

The per-second number matters because it changes the budgeting question. A firm does not buy “one demonstrative”; it usually buys a sequence of attempts: prompts, reference images, source video, client edits, associate revisions, partner comments, maybe a version with labels removed, maybe another version rebuilt after someone realizes the wrong scene angle implies too much. Under H3’s current public pricing structure, the price meter follows generated seconds, not the moment a trial team decides a clip is usable. [1]

That makes H3 different from the prepaid MiniMax Video Packages a procurement team may already have seen. Those packages currently run from a Standard plan at $1,000 per month for 3,760 points up to a Business plan at $6,000 per month for 26,780 points, but the same package page excludes H3. [2] For H3, there is no package buffer in the cited documentation, no prepaid hedge against a messy drafting cycle, and no easy way to treat experimentation as already absorbed by a subscription line item.

The legacy contrast is useful only because it shows the shift. MiniMax lists Hailuo-2.3 at $0.28 for a 768P 6-second generation, $0.49 for a 1080P 6-second generation, and Hailuo-2.3-Fast at $0.19 for a 768P 6-second generation. H3, by comparison, is priced by second at its output resolution. [1] That is simpler to calculate and easier to underestimate.

Where the hidden charges enter the legal workflow

The first hidden cost is not exotic: rejected output. If a 15-second 2K generation costs $1.95, five rejected attempts before one usable draft would put the generated-output spend at $11.70 for that one accepted clip, assuming all six generations are 15 seconds at 2K. That hypothetical is still pocket change compared with attorney time, but it is exactly how a “two-dollar clip” becomes a line item nobody can reconcile from the final presentation alone.

The second hidden cost is input material. MiniMax’s pay-as-you-go page states that the first five input images are free and that each additional image costs $0.04. It also states that input video is billed by duration at the output-resolution rate, while audio input is free. [1] For a legal team using references from site photos, deposition exhibits, accident-scene footage, medical animations, or day-in-the-life materials, that input rule matters almost as much as the output rate.

Take a clearly hypothetical production pass. A team asks for a 15-second 2K demonstrative, provides a short source video, uses more than five reference images, and then regenerates twice because the first version overstates an event sequence. The invoice is no longer just the final $1.95 clip. It can include generated seconds for each attempt, source-video duration billed at the 2K rate, and the additional-image charges. The dollars may still be small; the audit trail is not.

That audit trail belongs to someone. The associate asked to make a quick visual has to know which source file was used. The paralegal has to preserve the inputs and identify which images were merely references. Legal ops has to compare API billing against matter budgets. The lawyer who uses the clip has to be able to describe it without letting the technology’s polish imply factual certainty it does not have.

Feature claims are useful context, not admissibility evidence

H3 is worth attention because its claimed capabilities are directly relevant to demonstrative work: public early writeups describe 15-second generation, native 2K output, native audio, scene-persistent characters, and multi-reference conditioning. [3][4] Those are the kinds of features that can make a presentation editor lean forward, because they reduce the distance between a rough concept and a courtroom-ready-looking visual.

They do not answer the legal reliability question. Runware’s collection ranks MiniMax H3 as the No. 1 video AI model, but that is a product ranking, not a courtroom validation study. [5] As of this July 31, 2026 verification, the materials reviewed did not provide public H3 benchmark scores or a public model card that would let a legal buyer treat quality impressions as independent technical evidence.

The same caution applies to early price chatter. Kie.ai reported an early-tester claim of roughly $1 for a 15-second 2K clip, but that figure should be treated as unverified for budgeting unless and until MiniMax’s official documentation changes. [3] The controlling public math remains $0.13 per second at 2K, or $1.95 for 15 seconds. [1]

The risk is not that H3 has been sanctioned; it is that cheap video lowers the threshold for careless use

No sanction order in the materials reviewed names MiniMax H3. It is too new for that kind of record. The risk analysis is therefore category-level: photorealistic generative video used in legal presentation work now sits near documented failures involving AI-generated exhibits, metadata scrutiny, and lawyer duties that attach regardless of which model made the file.

Mendones v. Cushman & Wakefield, Inc. is the concrete warning. In that California Superior Court matter, the court imposed terminating sanctions on September 9, 2025, after finding misconduct involving GenAI deepfake exhibits identified as 6A and 6C. The order addressed metadata issues, including indications tied to an iPhone 6 and iOS 12.5.5, as part of the scrutiny of the challenged materials. [6] eDiscovery Today’s case-law writeup also treated the matter as a deepfake-video-and-image sanctions decision. [7]

That does not mean Mendones is an H3 case. It means a polished synthetic exhibit can become a sanctions problem when provenance, metadata, characterization, and disclosure fail. A two-dollar generation price does not reduce the duty to know what is being shown to a court.

Louisiana Act 250, effective August 1, 2025, adds a state-law example of the same pressure by imposing a reasonable-diligence verification duty for certain AI-generated legal filings and related uses. [8] ABA Formal Opinion 512, issued July 29, 2024, frames generative AI use through professional duties including competence, confidentiality, communication, candor, meritorious claims and contentions, fees, and supervision. [9] Those duties do not turn on whether the clip cost $1.35, $1.95, or $195.

The federal rulemaking picture is still unsettled. A proposed Federal Rule of Evidence 707 addressing machine-generated evidence had stalled at the Advisory Committee stage by May 2025, while some states moved ahead on their own AI-evidence and verification rules. [10] For a trial team, that makes local practice and judge-specific expectations more important, not less.

What to budget before anyone calls it a courtroom visual

For procurement, the H3 cost model should be separated into at least three buckets: generated output, input material, and legal handling. The first two are API math. The third is the expensive one.

• Generated output: seconds generated at the selected H3 resolution, including rejected drafts unless credited or reversed by the vendor.
• Input material: source video billed by duration at the output-resolution rate, additional images after the first five at $0.04 each, and audio input at no charge under the current MiniMax pay-as-you-go page. [1]
• Access structure: H3 pay-as-you-go only under current documentation, with prepaid Video Packages excluding H3 and 768P described as closed beta. [2][3][4]
• Legal handling: source collection, prompt and version logging, metadata preservation, attorney review, disclosure decisions, and a defensible explanation of whether the clip is demonstrative, illustrative, simulated, edited, or offered for some evidentiary purpose.

The last bucket is where the economics flip. A 15-second 2K generation can cost less than a cup of coffee, but the review chain may involve a paralegal, an associate, a partner, an expert, a client contact, and possibly a motion in limine. If opposing counsel challenges the clip, nobody will care that the API line item was small. They will ask what the clip depicts, how it was made, what materials conditioned it, what was changed, what was preserved, and whether the presentation misled the court.

A sensible internal rule is to price H3 experimentation cheaply but treat courtroom use expensively. Drafting a concept visual for internal strategy is one risk class. Showing a synthetic video to a mediator, judge, jury, client, regulator, or counterparty is another. The same file can move from one class to the other the moment someone exports it into a deck or exhibit binder.

The procurement answer

At official July 31, 2026 pricing, MiniMax H3 is inexpensive at the clip level: $1.35 for 15 seconds at 768P and $1.95 for 15 seconds at 2K. The meaningful buyer questions are whether the firm accepts pay-as-you-go-only exposure, whether it can actually access the 768P tier, how it will control retries and input-material charges, and who owns the verification record before a clip becomes a legal presentation asset.

For deeper risk context, keep the Mendones Risk Digest record near any discussion of synthetic video in court, read the ABA Formal Opinion 512 ethics record, and put any proposed use through a defensible AI workflow. This is a tool-evaluation cost article, not legal advice, and not a claim that H3 has already been tested in court.

References

1. Pay as You Go - Models - MiniMax API Docs, MiniMax API Docs.
2. Video Packages - Models - MiniMax API Docs, MiniMax API Docs.
3. What Is Hailuo H3? - Kie.ai, Kie.ai.
4. MiniMax H3 Review, Deevid.ai.
5. Best Video AI Models, Runware.
6. Mendones v. Cushman & Wakefield, Inc., No. 23CV028772, Superior Court of California, County of Alameda, September 9, 2025.
7. Deepfake Videos and Images Lead to Terminating Sanctions: eDiscovery Case Law, eDiscovery Today, September 25, 2025.
8. Act No. 250, Louisiana Legislature, 2025.
9. Formal Opinion 512: Generative Artificial Intelligence Tools, American Bar Association, July 29, 2024.
10. The Federal AI Evidence Rule Stalled, But Some States Didn’t Wait, Complete Legal.