MiniMax H3 Max API & Pricing: Endpoints, Costs and Setup

01

Choose the right H3 Max API route

fal currently exposes separate H3 Max routes for text-to-video, image-to-video and reference-to-video. Use text-to-video for prompt-only generation, image-to-video when a still should guide the opening composition, and reference-to-video when richer identity, style or audio conditioning is required.

The reference route is live in current fal documentation. Do not rely on older launch copy that described it as a future endpoint.

02

Keep authentication server-side

fal uses API-key authentication and its documentation recommends keeping `FAL_KEY` in a server environment. A browser or mobile client should call your authenticated backend rather than receive a persistent provider key.

Also apply your own user authorization and generation limits at that backend boundary so one leaked client session cannot create unlimited provider spend.

03

Use the asynchronous queue workflow

Video jobs are better handled as asynchronous work. fal documents submission that returns a request identifier, followed by status/result retrieval, with webhook support for long-running jobs.

1. Validate the prompt, media and requested settings.
2. Submit to the matching H3 Max endpoint.
3. Store the returned request ID.
4. Return control to the client quickly.
5. Poll status or receive a webhook.
6. Persist the result URL and cost-relevant metadata.

04

Important parameters and limits

The current text-to-video schema exposes prompt, duration, resolution, seed, safety checking, prompt-expansion behavior and aspect-ratio controls. H3 Max supports 480p or 768p output and clips from five to 15 seconds.

Image-to-video adds image inputs and can support first/last-frame workflows. Reference-to-video accepts richer conditioning media. Treat exact enum names and file limits as API implementation details and recheck the live schema before shipping a client library or strict validator.

05

Current H3 Max API pricing

The standard rates reverified September 7, 2026 are $0.05 per output second at 480p and $0.08 per output second at 768p. At those rates, five seconds costs $0.25/$0.40, ten seconds costs $0.50/$0.80 and 15 seconds costs $0.75/$1.20.

Reference inputs can add a second charge. fal's current guide says the first 4,096 reference tokens are free and additional reference tokens cost $0.02 per 1,000 tokens, with video-reference token usage depending on generated resolution and reference duration.

06

Control generation cost

• Draft at 480p and short duration while the brief is changing.
• Move selected outputs to 768p only when needed.
• Cap automatic retries.
• Track cost by job using duration, resolution and reference usage.
• Set per-user and per-workspace generation ceilings in your own application.

07

Set realistic latency expectations

fal reports under three seconds of backend inference for a five-second 768p clip on its optimized stack. That is not the same as guaranteed API wall-clock latency.

Uploads, prompt expansion, reference preprocessing, queue delay, network transfer and CDN delivery can all affect the time a user waits, so instrument your own end-to-end latency instead of repeating the backend figure as an SLA.

08

Production safeguards

Validate uploaded media and file types, preserve request IDs for debugging, rate-limit generation, protect keys, and apply your own rights/moderation checks for likenesses, copyrighted inputs, dialogue and branded material.

Seeds are useful for controlled comparison but do not make every output deterministic, especially when prompt expansion or other generation settings change.

09

Text vs image vs reference route

Text-to-video is the simplest route for rapid concept generation. Image-to-video is a better fit for product shots, artwork, characters or compositions that should begin from a known visual. Reference-to-video is the stronger choice when multiple source assets need to influence identity, style or audio.

If your actual requirement is 2K output or instruction-based video editing, standard MiniMax H3 may be a better model choice than H3 Max despite the latter's speed.

10

H3 Max API FAQ

Is H3 Max available by API? Yes. fal currently exposes text, image and reference-to-video routes.

Can I call it directly from a browser? A browser can send a request, but exposing a persistent API key client-side is unsafe; use a server-side proxy.

Does H3 Max support 2K? No. The current H3 Max routes target 480p and 768p; standard H3 is the broader option when 2K is required.

11

Read the main H3 Max guide

For the broader model overview, MiniMax Design availability, synchronized audio, speed context, limitations and H3 comparison, return to the main MiniMax H3 Max article.