MiniMax H3 Max, Turbo, and Director compared

Those workflows need different tests. A clip generator has to produce usable scenes fast enough to keep playback supplied. A session-based system also needs to respond to direction and preserve the story over time.

What the model names mean

MiniMax H3 is the broader multimodal model. MiniMax documents video generation, reference-based creation, and editing, with output options up to 2K.

H3 Max is the variant optimised for speed. MiniMax lists 480p and 768p output and clip durations of 5 to 15 seconds. Its documentation covers text-to-video and image-to-video workflows. fal also lists a separate reference-to-video endpoint, so check the specific provider and endpoint for the input you need.

H3 Max Turbo has separate text-to-video and image-to-video endpoints on fal. The text-to-video prices listed below are lower than Max's. Use the same prompts to compare the results before deciding whether Turbo meets the needs of your show.

H3 Max Director lets an application direct a continuous generation session. fal identifies it as the technology behind fal.live. You send direction during the session instead of managing a separate request for every finished clip.

How fast is H3 Max?

fal's August 2026 launch announcement reported generating a five-second video in roughly three seconds. The announcement also compared H3 Max with other models in fal's evaluation. These are vendor-reported results for that evaluation.

A "fastest AI video generator" claim leaves several questions unanswered for a livestream. How often does a request take longer? How many scenes need another attempt? How much time does processing and review add before a clip can play?

Measure the complete path to usable video with prompts from your planned show. Include scene transitions in the test. A model that produces an attractive isolated clip may still struggle to keep the same character recognisable across several generations.

H3 Max pricing on fal

Prices and Director access limits below were checked on September 6, 2026. Rates are in USD per generated video second. Max and Turbo rows refer to their text-to-video endpoints.

These rates come from fal's H3 Max, Turbo, and Director endpoint pages.

Director has a minimum charge of 60 seconds of session runtime. Its public page asks users to request access for sessions longer than two minutes.

Check the relevant endpoint before budgeting a run. Reference and editing workflows can have different prices, and the promotional rates above have stated expiry dates.

What does an hour of generated video cost?

For a clip-based estimate:

        Generation cost = billable video seconds × endpoint rate
      

At the listed H3 Max 480p promotional rate, 3,600 seconds of generated video costs $45. At the published post-promotion rate, it costs $180. This calculation assumes every generated second is used once at normal playback speed.

An AI livestream cost estimate also needs to account for discarded scenes and any additional billed attempts. Language-model calls, moderation, storage, and delivery add costs outside video generation. Replaying a clip can reduce the amount of new video required, while rejecting one can increase it.

In a September 2 post, Pieter Levels calculated $32,940 for 30.5 days of continuous generation at $0.0125 per second. That number is a rate-based estimate. It does not describe a monthly subscription or include every expense of running a public stream.

For a shared channel, separate the cost of generating the show from the cost of delivering it to viewers. A growing audience does not require a separate copy of each generation when everyone watches the same output.

How to use H3 Max for a first test

Open the relevant provider playground and start with a short scene. Choose text-to-video for a written description or image-to-video when you want to supply a starting frame. State the subject, action, setting, and camera movement in the prompt.

Here is an original prompt to try:

A small research robot stands beside a greenhouse window at sunrise. One leaf begins to move although the other plants remain still. The robot leans closer to inspect it. Keep the robot's appearance and the room layout consistent. Use one continuous shot with slow camera movement and quiet mechanical sounds.

This prompt has not been tested here. Use it to compare how models follow the same directions. For a second scene, change one event while keeping the robot and greenhouse consistent. Check whether details carry over when you supply the previous result as context through a supported input.

Compare Max and Turbo at the same resolution and duration. Record request time and billable output, then review whether the action happened and the subject stayed recognisable. Keep failed or unusable results in the comparison.

Test Director separately. Its documented realtime client uses an alpha API, and its session workflow differs from the clip endpoints. Check access limits before planning a longer trial.

Which should you test first?

For a clip-based prototype, we would compare Turbo and Max on a small set of scenes from the intended show. Start with Turbo's lower listed rate, then see whether Max produces enough additional usable material to justify its cost.

For a show built around ongoing audience direction, test Director's response to prompt changes and its behaviour when a session ends. Confirm that your account can run the session length you need.

For reference-based creation or editing, check the relevant H3 and provider endpoints before choosing a model. Support for one workflow does not establish support for another.