MiniMax Hailuo 3 Joins Video Lab — Native-Audio 2K, And It Goes First

August 9, 2026

MiniMax Hailuo 3 is in Video Lab: native-audio clips at 2K for commercial creative, product shots, readable text, and controlled cinematic motion. It is now the first provider Auto tries, it accepts reference images, and it bills by clip length instead of one flat rate. Higgsfield is no longer a video provider (Higgsfield Soul in Image Lab is unchanged).

What you can do

• Generate video from text — describe the subject, action, camera movement, lighting, and sound; Hailuo 3 renders the complete clip at 2K (2560×1440 for 16:9), 5 to 15 seconds long.
• Animate an existing image — use a session image as the first frame, with an optional last frame for a controlled transition.
• Pass up to nine reference images — reference_images drives reference-to-video, alone or alongside first/last frame anchors.
• Get native audio with the picture — synchronized dialogue, music, ambience, and sound effects without a separate audio pass. Toggle it with minimax_audio (auto / on / off), and optionally stamp the AIGC watermark.
• Use six output shapes — 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16 cover cinematic, standard, square, and vertical work.
• Let Auto reach for it first — the chain is now Hailuo 3, then xAI Imagine, then Runway. Pin provider: "minimax" when you want Hailuo 3 specifically.
• See what you were charged — every completed Video Lab result carries charged_tier and charged_tokens.
• Inspect audio before editing — Video Lab detects embedded audio and adds a storyboard summary, so you can tell whether subtitles, narration, or another soundtrack are actually needed.

Where this shows up

You need a product hero clip with readable branding, controlled rotation, studio lighting, and matching sound. Pin MiniMax Hailuo 3 instead of assembling separate video and audio generations.

You already have an approved key visual. Animate it through image-to-video while preserving the subject, composition, and brand treatment.

You upload a cast still plus wardrobe references. Hailuo 3 takes those as reference_images rather than only a first frame.

You ask for a product hero with native audio on Auto. It goes to Hailuo 3 immediately instead of burning a cheaper draft and only reaching MiniMax after two failures.

Try it

• "Use MiniMax Hailuo 3 to create a 5-second 16:9 product film: a white ceramic coffee cup marked ALFRADA rotates slowly on a clean studio pedestal, soft natural shadows, subtle camera push-in, quiet room ambience."
• "Generate an 8-second 2K café scene with native audio — barista calling out an order, ambient chatter. Use Video Lab auto."
• "Create a 6-second Hailuo 3 clip using these three reference stills for the character's face and wardrobe."

Heads up

• Pricing follows duration. MiniMax 5–8s ≈ 5M tokens, 9–12s ≈ 8M, 13–15s ≈ 10M, plus 200k per reference image. xAI is 3M (1–8s) or 5.5M (9–15s). Runway stays on the previous flat 2.5M until we publish Runway SKU tiers. enhance_prompt is near-free.
• OpenRouter Hailuo 3 only sells 2K. A 720p or 1080p request is upgraded and billed at 2K; minimax_resolution: "auto" and "2K" mean the same thing.
• Supported duration is 5–15 seconds. Requests outside that range are adjusted and surfaced in the result.
• Image-to-video needs a session image. Use input_image; add last_frame_image only when you want a defined ending composition.
• Higgsfield is not a video option anymore. Image Lab still uses Higgsfield Soul when Gemini refuses.
• Native audio may already contain dialogue, ambience, or music. Check has_audio and the storyboard before layering another track in Video Editor.