Back to models
Model · MiniMax
MiniMax

MiniMax H3 Max

minimax/minimax-h3-max
Try in Studio

The default: cheap, any length 5–15 s, native stereo sound. Strongest at many references at once (9 images, 3 clips, 3 audio, each given a job in the prompt), legible on-screen text and UI, and localized edits of an existing clip passed as Video 1. Also end frames, first/middle/last keyframes, extend, the explicit 3D camera path, inserting a scene, previs to footage, and recasting people from photos.

image to videoreference imagesnative audiofirst and last framerefframesmidframeextend a clipcamera controlinsert a scene3d render to videorecast peoplelong clips
Specs
Provider
MiniMax
Kind
Video
Pricing
$0.064–$0.576 / sec
Endpoint
/api/v1/studio/video/generate
Providers and pricing
OrderProviderList priceOur priceNotes
1Vercel AI Gateway$0.050–$0.080 / sec$0.064–$0.102 / sec2 tiers
2fal$0.050–$0.450 / sec$0.064–$0.576 / sec3 tiers
3Atlas Cloud$0.050–$0.205 / sec$0.064–$0.262 / sec4 tiers
TierVercel AI GatewayfalAtlas Cloud
480P$0.064list $0.050$0.064list $0.050$0.064list $0.050
768P$0.102list $0.080$0.102list $0.080$0.102list $0.080
1080P—$0.205list $0.160—
extend-2K—$0.410list $0.320—
insert-480P—$0.064list $0.050—
insert-768P—$0.077list $0.060—
recast-768P—$0.384list $0.300—
recast-1080P—$0.576list $0.450—
1440p-sr——$0.192list $0.150
4k-sr——$0.262list $0.205

Each request goes to the cheapest provider that can run it at the tier you ask for (highlighted); on a tie, Vercel AI Gateway first, then fal, then Atlas Cloud. If that provider fails on its side, the next one takes the request. You pay the list price of the provider that ran it, plus 28%.

Live performance
Loading measurements…
How we measure. Every request is timed from the moment we send it. A finish time comes from the provider's own timestamps when it reports them, from the moment it calls us back when it doesn't, and otherwise from polling: we take the midpoint between the last check that found the job running and the first that found it done, so it is off by at most half the poll interval. Requests you stop are left out. Figures appear once a window has 20 completed requests, and are aggregate only.
Code
curl https://terramind.com/api/v1/studio/video/generate \
  -H "Authorization: Bearer $TERRAMIND_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3-max",
    "prompt": "A cinematic shot of a sailboat at sunset"
  }'

Want to run this in your app? Grab an API key · Video API reference

More by MiniMax
ModelContextInOutCapabilities
MiniMaxminimax/minimax-m2.7205K$0.300$1.20
caching
MiniMaxminimax/minimax-m2.7-highspeed205K$0.600$2.40
caching
MiniMaxminimax/minimax-m31M$0.300$1.20
reasoningcaching
MiniMaxminimax/minimax-m3.1-flash1M$0.150$0.600
reasoningvisioncaching
MiniMaxminimax/minimax-h3-max-turbo0$0$0
image to videonative audiofirst and last frame