MiniMax H3 Adds Video Edit, Video Extend, and a 540p Tier

Since the open-weights MiniMax H3 launched on WaveSpeedAI in early August, the family has grown from three endpoints to eight, and this week it gained a new resolution tier. This post is the reference for what was added, what each endpoint takes, and what it costs. Every figure is from the live catalog on September 5, 2026.

If you are new to the model, MiniMax H3 open weights vs the official API explains why the WaveSpeedAI endpoints exist at all, and MiniMax H3 resolution covers the 768p native canvas.

What was added

Endpoint	Added	Input	Base rate (480p, per second)
wavespeed-ai/minimax-h3/video-edit	September 2	video + prompt, optional references	$0.05 per counted second
wavespeed-ai/minimax-h3/video-extend	September 2	video + prompt, optional last frame	$0.04
.../text-to-video-lora, .../image-to-video-lora, .../reference-to-video-lora	August 16	same as the base endpoint plus up to 3 LoRAs	$0.05 to $0.06
540p tier on every endpoint above and the three originals	September 4	resolution: 540p	1.5x the 480p rate

The three original endpoints, text-to-video, image-to-video, and reference-to-video, are unchanged apart from the new tier.

Video edit

wavespeed-ai/minimax-h3/video-edit is the one people have asked for most. You pass a video URL and a prompt describing the change; the source drives subject identity, composition, and motion, and the model rewrites what the prompt names: lighting, weather, style, environment, wardrobe, or a specific object.

Parameters that matter:

• reference_images (up to 9) and reference_audios (up to 3) guide the edit. Refer to them in the prompt as <Picture 1> and <Audio 1>.
• generate_audio defaults to true and produces a new native soundtrack. Set it to false to keep the input video’s original audio on the edited output.
• duration (3 to 15) defaults to the input length, rounded up, with a 3-second floor. aspect_ratio defaults to the input’s.
• resolution: 480p, 540p, 768p, or 1080p.

Billing counts input seconds plus output seconds, each capped at 15, times the tier rate, plus $0.02 per reference image or audio. A 5-second clip edited into a 5-second output at 480p is 10 counted seconds at $0.05, so $0.50. At 540p the rate is $0.075 per counted second ($0.75), at 768p $0.125 ($1.25), at 1080p $0.25 ($2.50). Input duration is billed capped at 15 seconds.

Use it for restyling a clip you already like, fixing a scene’s time of day or weather, or swapping a product in a shot while keeping the camera move. It is not a face-swap or a cut tool; if you want to change who is in the shot rather than what surrounds them, reference-to-video with a reference image is the better fit.

Video extend

wavespeed-ai/minimax-h3/video-extend takes a video and a prompt for what happens next, generates a new segment from the last frame, and returns the original plus the new segment concatenated. last_image optionally pins where the extension should end, and audio is generated for the new segment.

Only the appended segment is billed, at the image-to-video rate: $0.04 per second at 480p, $0.06 at 540p, $0.08 at 768p, $0.16 at 1080p, for 3 to 15 seconds per call. The original footage is free. Chaining calls is the way to pass H3’s 15-second single-shot limit: a 5-second base clip extended twice by 10 seconds each is 25 seconds of video for the price of 20 generated seconds.

Match resolution to the source. Extending a 768p clip at 480p produces a visible quality step at the seam.

The 540p tier

Every public H3 endpoint now accepts resolution: 540p, priced at 1.5x the 480p rate. Per generated second on text-to-video, image-to-video, and video-extend:

Resolution	Per second	5 s	15 s
480p	$0.04	$0.20	$0.60
540p	$0.06	$0.30	$0.90
768p (native)	$0.08	$0.40	$1.20
1080p	$0.16	$0.80	$2.40

Reference-to-video and video-edit follow the same 1 / 1.5 / 2.5 / 5 ladder from a $0.05 base, and the LoRA variants from $0.05 (text and image) or $0.06 (reference).

The tier exists because 480p was doing two jobs: cheap previews, and final output for surfaces such as chat thumbnails and in-feed autoplay where 768p is more than needed. 540p gives that second group a sharper frame for a 50 percent premium instead of a 100 percent one, while the 768p native canvas remains the tier where the model was trained.

Which endpoint for which job

• Text or single image to a finished clip: text-to-video or image-to-video at 768p. Use 480p or 540p for iteration, then re-run the same seed at 768p.
• Keep a real person, product, or style consistent across clips: reference-to-video with up to 9 reference images.
• Change the look of footage you already have: video-edit.
• Make an existing clip longer: video-extend, chained if needed.
• Your own style or character adapter: the LoRA variants, up to 3 LoRAs per call.
• Full HD delivery: resolution: 1080p on any endpoint.

All eight endpoints share one API shape: POST https://api.wavespeed.ai/api/v3/<model_uuid> with the JSON body, then poll GET /api/v3/predictions/<id>/result. Prompts written for one carry to the others, which is the point of keeping them in one family rather than routing to different vendors.

Start from the MiniMax H3 API page, where each endpoint’s playground shows the live price for the exact parameters before you submit.