Hailuo and MiniMax H3 for AI Micro Drama: Versions, Pricing and a 9:16 Workflow

The short version

Use MiniMax H3 for talking scenes, and Hailuo 2.3 for silent action shots. "Hailuo" is MiniMax's video brand and app. For a long time it meant Hailuo 02 and Hailuo 2.3, which are fast, cheap, great with bodies and faces, and completely silent.

On July 31, 2026, MiniMax released H3. It makes the picture and native stereo audio together. It can also cut between shots inside one clip.

H3 takes up to 9 reference images (pictures you give it so it copies a face or place).

The Hailuo app calls it hailuo3.0, and the API spec calls it Hailuo-03. It's the same model.

That's a big shift for micro drama. Hailuo 2.3 is still good for silent action, reactions and establishing shots. But H3 is the one to build a talking vertical series on.

MiniMax's own feature gallery shows a 9:16 vampire-romance trailer as its story example. It's built to have a "ReelShort / DramaBox feel."

Which Hailuo is which

The takeaway: only the two H3 models make sound, and they're also the ones with longer clips.

MiniMax's pricing page now lists Hailuo 2.3, 2.3 Fast and 02 under "Legacy Models." They still work, but start new projects on H3.

Good and bad for micro drama

Good:

• Bodies and faces. MiniMax sold Hailuo 2.3 on body movement, micro-expressions (tiny face movements) and physics. That's what slap reveals and stare-downs need.
• Steady look. One creator's tests found it keeps your start frame's color and lighting.
• Talking scenes in one go (H3). Dialogue comes out with the picture. MiniMax says it's stable in 11 languages, including English, Spanish and Portuguese, which helps with dubbed versions.
• Cuts inside a clip (H3). Write a 15-second beat as three timed shots. The same people and room carry across the cuts.
• Price. H3 at 768P costs $0.08 a second on the MiniMax API and $0.06 on fal.

Not so good:

• Hailuo 2.3 is limited. It has no audio and no aspect-ratio setting for text-to-video. So vertical shots must start from a 9:16 still (image-to-video: you give it a picture and it animates it).
• Short clips. Nothing runs past 15 seconds per request. A 90-second episode is still 8 to 20 generations.
• One or the other on H3. It won't take reference images and a first or last frame in the same request. For each shot, you pick identity (references) or continuity (frames).

Where to run it

• Hailuo app (hailuoai.video) is the easiest place to test. On October 1, the create panel charged 7 credits per second for H3 at 768p and 12 at 2K. H3 Max at 480p cost 4.
• MiniMax API: an API means you send requests from code instead of using the app. H3 needs pay-as-you-go. The prepaid Video Packages only cover the older Hailuo models.
• fal.ai has H3, H3 Max and Hailuo 2.3. It also has extras like H3 Max Lip Sync, video extend and H3 LoRA training. A LoRA is a small add-on you train on your character's face.
• Replicate runs minimax/h3 at MiniMax's per-second rates.
• Self-hosting: the H3 weights are on Hugging Face, with ComfyUI templates. But the community license excludes the US, EU, UK and South Korea, so you'd apply for a separate license there. See self-hosting and ComfyUI workflows.

Pricing (as of October 2026)

The takeaway: with sound, 480p starts at $0.05 a second, and H3 on fal is the cheapest at 768p ($0.06).

What an episode costs. Take a 90-second episode and assume three takes for every keeper. That's our assumption, so plug in your own hit rate.

Drafting on H3 768P with the MiniMax API costs 270 seconds × $0.08 = $21.60. Regenerating only the keepers at 2K adds 90 × $0.05 = $4.50. The same drafts on H3 Max at 480P cost $13.50.

Silent Hailuo 2.3 at 768P costs about $12.60. That's before you pay for voices and lip sync (mouth movements that match the words).

The app goes less far than it looks. 1,000 Standard credits buy about 142 seconds of H3 at 768p, which is less than one episode with retakes.

So put series work on the API. The cost breakdown compares other models.

Best practices for 9:16 serialized drama

Lock the cast before episode one

Make a reference pack for each recurring character. Use a clean front view and a three-quarter view in neutral light, in their main costume. Add one photo for each main location.

On H3, pass them as reference images and name them in order in the prompt ("Image 1 is Adrian Vale, Image 2 is Elena Brooks"). Repeat the same physical description word for word every time. H3 also takes up to 3 voice reference clips, 2 to 15 seconds each.

Here's a creator tip worth stealing. Record voice references with different sentences than your script, so you can tell matching from copying.

For a 60-episode run, price out fal's H3 LoRA trainer. It's about $10 for 1,000 steps on the image-to-video trainer. There's more in character consistency.

Get vertical right

• H3 text-to-video: set the ratio to 9:16. You have to, since adaptive isn't allowed in text-only mode.
• H3 image-to-video: the output follows your image, so crop the still to exactly 9:16.
• Hailuo 2.3: there's no ratio setting. Always go image-to-video from a 9:16 still (first frames can range from 2:5 to 5:2).

For phone-screen framing tips, see vertical framing and shot language.

Speak each model's camera language

Hailuo 2.3 reads 15 bracket commands. They include [Push in], [Pull out], [Pan left], [Truck right], [Pedestal up], [Tilt down], [Zoom in], [Shake], [Tracking shot] and [Static shot].

You can combine up to three in one bracket ([Truck left,Pedestal up]) or chain them in order. MiniMax says exact commands work better than free text. Prompts can be up to 2,000 characters.

H3 wants a timeline. MiniMax's guide splits a prompt into integrated_multimodal_description, overall_soundscape and non_diegetic_music. Shots run [Shot 1], then [Shot 2] At 00:05.000, the camera cuts to....

Write camera moves as type plus amplitude plus speed ("the camera pushes in with small amplitude at slow speed"). Amplitude just means how big the move is. MiniMax's h3-prompt-writing skill on GitHub turns rough notes into this format.

Dialogue and lip sync

On H3, give each speaker an ID and wrap the line: (S1) says: <d>[English] Sign it.</d>. For narration, call it an off-screen voiceover. Then say the on-screen character's lips stay closed.

Let one person speak at a time, and tell the listener to keep their mouth shut. That's how one creator's test with two characters and two voices kept the voices from swapping.

Hailuo 2.3 can't talk. Make the lines with TTS (text-to-speech), then add lip sync. MiniMax Speech 2.8 HD costs $100 per million characters, and a rapid voice clone costs $1.50.

Watch out: clones you don't use for 7 days get deleted. On fal, H3 Max Lip Sync animates a still from audio at $0.08 a second (768p). For the full workflow, see lip-sync and dubbing.

Clip length, stitching and pacing

Micro drama runs on a 3-second hook and a cliffhanger every minute or so. Plan episodes as beats of 5 to 15 seconds. Use H3's in-clip cuts for back-and-forth dialogue, and single clips for inserts and reactions.

To carry a scene across generations, use the last frame of one clip as the next first frame. That works on Hailuo 2.3 too.

Some creators lay down a "sound skeleton" of music hits before editing, so cuts land on the beat. A slight speed-up in the edit can also fix floaty motion. See editing and captions.

Common mistakes

• Getting horizontal video from Hailuo 2.3 text-to-video. There's no ratio control. Start from a vertical still.
• Asking for 1080P at 10 seconds. Hailuo 2.3 only does 1080P at 6 seconds.
• Mixing references with keyframes on H3. The API rejects it, so split the shot.
• Letting prompt expansion rewrite your blocking. Prompt expansion means the model rewrites your prompt. H3 Max and fal's H3 let you set it to disabled.
• Budgeting on promo prices. fal's H3 Max launch discount ended September 30.
• Real faces and brands. MiniMax's moderation may block content that looks like it uses other people's rights. Keep your cast original.

Prompts and next steps

Grab 40 ready-to-paste Hailuo and H3 prompts. Or jump to a genre: CEO romance, revenge, werewolf, historical, thriller, comedy, fantasy, family.

Still deciding? Read choosing a video model, then follow how to make an AI micro drama.