fal H3 Max: API, Speed, Pricing, and H3 Differences (2026)

fal H3 Max is a post-trained MiniMax H3 variant built for faster 480p and 768p audiovisual generation. fal announced it on August 27, 2026, with text-to-video and image-to-video API endpoints, native synchronized audio, 5–15 second durations, and an optional final keyframe. fal reports that a 5-second 768p clip renders in under 3 seconds; that is an inference claim, not a promise of identical end-to-end latency under every queue and network condition.

The launch is meaningful for iterative production: a model that can return a draft faster than real time changes how many camera, motion, and dialogue options a team can review. It does not replace standard MiniMax H3 in every workflow. H3 Max stops at 768p at launch, while standard H3 offers 2K output, multimodal reference-to-video, and video editing.

fal H3 Max Quick Facts

What Is H3 Max?

H3 Max is fal's hosted, post-trained version of the open-weight MiniMax H3 base model. fal says it introduced additional training data and used its reinforcement-learning framework to improve prompt adherence, audiovisual quality, and aesthetics. The company also co-designed the serving path around its inference engine instead of deploying the model through a generic stack.

The distinction matters. MiniMax created the H3 foundation model; fal created and serves the H3 Max variant. The launch materials announce fal API access, not downloadable H3 Max weights. Do not infer that the post-trained H3 Max checkpoint has the same release status or license as the underlying MiniMax H3 weights unless fal publishes those terms separately.

What Is the H3 Max News?

fal's August 27 announcement makes three claims worth testing:

1. Speed: fal reports roughly 2.5 seconds of backend denoising time and under 3 seconds of inference for a 5-second 768p clip.
2. Preference: fal's internal human evaluation placed H3 Max first for overall preference, prompt understanding, and aesthetics against twelve named video models.
3. Availability: the model launched on fal's Playground, Agent, and API, with a temporary 50% API discount and a daily free sandbox allowance.

The speed number describes model inference. Real application latency also includes prompt expansion, queue time, uploads, downloads, webhook handling, and retries. fal's quality prompt-expansion mode alone may spend up to about 30 seconds rewriting a prompt, while balanced generally adds about one second.

Does Independent Benchmarking Support the Quality Claim?

Independent preference data supports the claim within a specific test category, but not every production use case. On August 28, the Artificial Analysis image-to-video leaderboard with audio listed MiniMax H3 Max first with an Elo score of 1,202, a ±9 confidence interval, and 5,322 samples. The live table changes as new votes arrive.

Design Arena's image-to-video board also placed H3 Max first in the launch snapshot cited by fal. fal's own benchmark used head-to-head human ratings for overall preference, prompt understanding, and aesthetics, aggregated with Bayesian Elo.

These leaderboards answer “which output did voters prefer for the tested prompts?” They do not establish character consistency across a full episode, editability after approval, queue stability, safety-filter behavior, or cost per accepted shot. A production decision still needs a test set drawn from your own scripts and reference assets.

H3 Max API Inputs and Outputs

The launch exposes two provider-specific model IDs:

Both endpoints return a video URL. Responses can also include the expanded prompt and timings.inference, which records backend denoising time rather than total request time. The image endpoint follows the first image's aspect ratio; the text endpoint lets you select six documented ratios.

fal documents three prompt-expansion modes: disabled, balanced, and quality. Start with balanced when turnaround matters. Use quality only when the richer rewrite justifies up to roughly 30 seconds of extra latency, and retain the returned expanded prompt so an approved shot can be audited later.

Minimal JavaScript Request

Keep FAL_KEY on the server and use the current @fal-ai/client package:

import { fal } from "@fal-ai/client";

const result = await fal.subscribe("minimax/h3-max/text-to-video", {
  input: {
    prompt:
      "Five-second medium shot of a cel-shaded detective entering a rainy alley. Slow dolly in. Coat and hair move in the wind. Sound: rain, distant traffic, one measured footstep.",
    duration: 5,
    resolution: "768P",
    aspect_ratio: "16:9",
    prompt_expansion_mode: "balanced",
  },
});

console.log(result.data.video.url);
console.log(result.data.timings?.inference);

For production traffic, fal recommends queue submission and webhooks instead of holding a request open. Store the model ID, request ID, original prompt, expanded prompt, seed, price snapshot, output URL, and measured wall-clock latency with each result.

How Much Does H3 Max Cost?

As checked on August 28, the live text-to-video endpoint lists a launch price of $0.025 per output second at 480p and $0.04 per output second at 768p. It says the promotion ends September 1, after which the regular rates are $0.05/sec and $0.08/sec respectively.

fal's H3 Max landing page describes the discount as lasting fourteen days, while the endpoint names September 1. Because these two official pages conflict, treat the endpoint's live price as operational truth at submission time. The landing page also says signed-in users receive five free generations every rolling 24 hours, up to 15 seconds at 768p; trial terms can change without an API version change.

The generated-second price is not the same as production cost. Budget for rejected generations, prompt-expansion time, uploads, storage, review, edits, and final upscaling. Compare models by cost per approved second, not the cheapest first render.

H3 Max vs MiniMax H3

Choose H3 Max for fast draft loops, dialogue or sound-aware previews, camera tests, and high-volume 768p generation. Choose standard H3 when 2K delivery, video or audio references, multi-asset identity control, or localized video editing is essential. A hybrid workflow can draft on H3 Max and reserve H3 or another high-resolution model for selected final shots—but verify that regenerated shots preserve the approved motion and timing.

How to Test H3 Max for Motion Comics

Do not benchmark with one attractive prompt. Build a short sequence that reflects the failures your audience will notice:

1. Dialogue close-up: one character speaks a short line with a defined emotion and quiet room tone.
2. Repeat close-up: the same character returns with unchanged face, costume, palette, and voice direction.
3. Two-character shot: preserve blocking, eyelines, scale, and speaking order.
4. Full-body action: include hands touching a prop, a camera move, and a clear start and end pose.
5. Vertical variant: repeat an approved shot at 9:16 without losing the subject or key action.
6. First-to-last-frame transition: test whether the path between two approved storyboards remains physically and narratively plausible.

Run each case with fixed evaluation criteria: instruction adherence, identity, motion, composition, lip sync, sound, artifacts, latency, and accepted-shot cost. Repeat enough seeds to expose variance. Leaderboard rank can choose what to test first; only this sequence test can decide whether H3 Max belongs in your episode pipeline.

Where AniKuku Fits

H3 Max solves shot generation. It does not manage the script, break a scene into shots, preserve approved character references, or track which render belongs in an episode. AniKuku's production workspace keeps scripts, shot lists, character and scene assets, storyboards, and output versions together before and after an external model call.

That separation makes model testing safer. Prepare the same shot package once, render selected shots with H3 Max and comparison models, then review outputs against the same brief. If a better model appears, the story structure and approved assets remain usable instead of being trapped in one provider's generation history.

Frequently Asked Questions

Is H3 Max made by MiniMax or fal?

fal created H3 Max by post-training MiniMax's H3 base model and optimizing it with fal's inference stack. MiniMax made the underlying H3 foundation model. Use minimax/h3-max/... when referring to fal's hosted provider IDs.

Is H3 Max open source or open weights?

The underlying MiniMax H3 base has released weights, but fal's H3 Max announcement provides hosted API access and does not announce a downloadable H3 Max checkpoint. Treat H3 Max as a fal-hosted model unless fal publishes separate weights and license terms.

Does H3 Max generate audio?

Yes. fal documents native synchronized audio for H3 Max, including dialogue, music, effects, and ambience described in the prompt. Audio quality and lip sync should still be tested in the language, voice style, and shot length you plan to ship.

Does H3 Max support 2K?

No at launch. H3 Max supports 480p and 768p. fal directs users who need 2K, reference-to-video, or video editing to standard MiniMax H3.

How fast is H3 Max?

fal reports under 3 seconds of inference for a 5-second 768p clip, with an example response showing about 2.53 seconds of backend denoising. Total user-visible time can be longer because of prompt expansion, queues, transfers, and application processing.

What is the best H3 Max setting to start with?

Start at 768p, 5–10 seconds, with balanced prompt expansion. This matches fal's recommended quality and latency starting point. Then test 480p for cheaper drafts or quality expansion when richer prompt interpretation is worth the extra wait.

Sources and Verification Date

This article was checked on August 28, 2026. Prices, free allowances, rankings, and available endpoints can change.

• fal H3 Max launch announcement
• fal H3 Max product page, free allowance, benchmarks, and FAQ
• fal H3 Max text-to-video endpoint and live pricing
• fal H3 Max image-to-video API schema
• Artificial Analysis image-to-video leaderboard with audio
• Design Arena image-to-video leaderboard
• fal standard MiniMax H3 capabilities