MiniMax H3 2k pricing: API Cost Breakdown & Guide

MiniMax H3 2k pricing is built around output duration rather than a flat subscription fee. The documented launch rate is $0.13 per second for 2K video, while 768p output is listed at $0.09 per second. This makes the model easier to estimate for individual clips, advertisements, product concepts, and short cinematic tests.

The listed price covers the generated output, but a heavily referenced request can add input charges. Reference images, reference videos, and output duration should therefore be calculated together before a production batch begins.

Video Highlights:

• H3 supports text, image, video, and audio references in one generation context.
• 2K output is priced at $0.13 per second.
• A 15-second 2K result has a listed base cost of $1.95.
• Commercial examples include product websites, fashion ads, animated posters, and film openings.

The 2K rate is the central figure for creators who need higher-resolution output. A short five-second test costs less than one dollar at the listed rate, but repeated generations can quickly become the larger budget factor. For that reason, use lower-cost tests where the workflow allows it, then reserve 2K generations for approved prompts and final candidates.

MiniMax H3 is designed around what the launch material calls an omni-reference workflow. A request can combine a written prompt with reference images, reference videos, and audio. This is useful when the goal is to preserve a character identity, imitate a camera movement, maintain a brand look, or align a soundtrack with a generated scene.

The output rate is only one part of the calculation. The first five reference images are listed as free, while additional reference images cost $0.04 each. Reference audio is listed as free. Reference video is billed according to its input duration at the selected output rate, so a request with several video references may cost more than the output alone.

A practical cost formula is:

Estimated request cost = output seconds × output rate + billable reference charges

For example, a 15-second 2K clip has a $1.95 base output cost. If the request uses seven reference images, the first five are free and two additional images add $0.08. The estimated total becomes $2.03, excluding any billable reference-video duration.

These examples show why reference selection matters. More inputs can improve creative control, but they also make the request more expensive when image or video thresholds are crossed. Start with only the references that solve a specific continuity problem.

The API is the operational path for MiniMax H3 in the current launch phase. The model accepts a required text prompt and supports several reference modes. A request can use reference media, or it can use a first frame, a last frame, or both. These modes should not be mixed: first-and-last-frame generation cannot be combined with the reference image, video, and audio mode.

The documented limits are important for production planning. H3 supports up to nine reference images, three reference videos with a combined duration of 15 seconds, and three audio references. The available duration is generally described as five to 15 seconds in the model overview, while the create endpoint lists four to 15 seconds. Treat the active endpoint documentation as the practical source when implementing an integration.

MiniMax H3 is positioned as a commercial video generator rather than only a cinematic experiment. The documented examples emphasize tasks that expose weaknesses quickly, including readable titles, product presentation, vertical poster design, fashion advertising, and branded web concepts.

Those examples are useful indicators of intended capability, but they should not be treated as a guarantee of repeatable production performance. A polished demonstration confirms what the model can produce in selected cases; it does not reveal failure frequency, rejected generations, latency, or identity consistency across a complete campaign.

An independent arena signal cited in the launch discussion placed H3 second in text-to-video with audio, with an Elo score of 1242 from just over 5,000 samples. Gemini Omni Flash was listed first at 1246. Because the reported 95% confidence ranges overlap, the cautious interpretation is near parity in that arena rather than a decisive lead.

The launch status also has an important distinction. The hosted API is available, while the model weights are not yet available for download in the current material. The expected weight release remains subject to applicable laws and regulations. This means H3 should currently be evaluated as an API-first service, not as a self-hosted open model.

Documentation can change quickly during an API launch. The model overview describes clips as five to 15 seconds, while the create endpoint lists four to 15 seconds. The pricing material lists 768p, but the active create endpoint described in the launch discussion exposes only 2K as an available resolution.

For that reason, use the endpoint documentation as the operational reference when building requests. Pricing pages are still useful for estimates, but implementation decisions should follow the fields and limits accepted by the live API.

The official MiniMax website is the appropriate place to check current product announcements and service information. Reconfirm rates, output options, authentication requirements, and model availability before committing to a recurring production budget.

Do not assume that a listed per-second rate represents the full cost of a campaign. Account for retries, prompt experiments, reference media, storage, post-production, and any workflow-specific API charges. A small test budget can reveal the model’s consistency before a larger batch is submitted.