minimax-h3-768p and minimax-h3-2k use the same video task API. The create call returns a task object, not the finished video file.
Poll and download:
Send JSON. Put every field at the top level. Do not wrap these fields in metadata. Image inputs must be public URLs that stay available until the task finishes. These models are not the 1440p minimax-h3 model. Do not send start_frame, end_frame, or reference_images here.

Models and prices

Billing is per generated second: duration × unit price. Failed tasks are refunded in full. An 8-second clip costs ¥2.80 Flex or ¥4.00 Official at 768p, and ¥4.48 Flex or ¥6.40 Official at 2K. The official route is the fastest, about 30 seconds per video. The 933 route passes face checks. Audio references are free. Reference video adds ¥0.18 per second. Each image after the fifth adds ¥0.10. H3 Context IR, for stronger multimodal context, costs ¥0.1 per call.

Parameters

Image-to-video

Send only image_url. Do not also send first and last frame fields.
For 2K, change model to minimax-h3-2k.

First and last frames

Send only start_image_url and end_image_url. Do not send image_url. Do not put the last frame in extra_images, and do not expect @Image2 to lock the last frame. Describe a transition from the first image to the last image. Use 8 seconds, and match the aspect ratio to the images.
If the two images differ a lot, a completed task does not guarantee both ends stay locked. Check the opening and ending frames before relying on the result.

Polling

Poll every 3 to 5 seconds. Stop when status is completed.
Read url from the response. That address downloads the video file directly. If status is failed, read error and do not recreate the same request indefinitely.