MiniMax H3 first request: a 5-second 768p clip for 38 cents on Sume

The cheapest useful first request to MiniMax H3 on Sume is 5 seconds at 768p: $0.38, with stereo sound included. H3 runs 5 to 15 seconds at 480p or 768p natively and prices 2K and 4K as upscales, so $0.32 gets you a 480p draft. Magic Hour's tracker, read 2026-10-06, dates MiniMax H3 to 2026-07-31 and lists it at 15 seconds, 2K, with native stereo audio, and open weights on 2026-08-03.

The vendor facts are on the tracker. The Sume limits and rates are in the Video Router docs and video docs.

Price per clip, read 2026-10-06

Billed price is the provider list times 1.25. The upscale rows are priced per second at the upscaled size.

The request

No sound flag is needed, because H3 stereo audio is always on, so there is no toggle to send.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/videos",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "minimax-h3",
        "prompt": "Rain on a neon street, a taxi passes, wet tires hiss",
        "duration": 5,
        "resolution": "768p",
        "aspect_ratio": "16:9",
    },
    timeout=60,
)
print(r.status_code, r.json())

After it runs

The reply is a 202 with an id and a polling URL. Poll until the status is completed, then download the file from unsigned_urls[0]. Listen for the sound first: with stereo audio always on, the prompt is how you steer it, so name the sounds you want in the text.

Then decide. If 768p is enough, extend the length to 8 or 10 seconds. If you want 2K, ask for it as an upscale and expect the higher per-second rate in the table. A 5-second 4K clip is $1.00, which is more than twice the 768p price, so use it for the final render only.

What 15 seconds costs, and when to go there

The full 15-second range at 768p is $1.13 and at 480p it is $0.94. For comparison, a 15-second Kling 3 clip with sound is $3.15. Go to 15 seconds when the scene needs a whole beat, such as a person walking in, speaking and leaving. For atmosphere shots, 5 to 8 seconds is enough, and the cost scales in a straight line with the number of seconds.

Reference inputs and lip sync

H3 accepts reference-to-video requests; the first five reference images add no fee and each image after that adds $0.08 at list (about $0.10 billed). For a talking clip from a still, H3 Max has its own lip-sync endpoint. Neither is needed for a first test, but both are the reason to pick H3 over a plain text-to-video model later. Start with the plain 5-second run, and add a reference only when the first result shows what is missing.

Open weights exist for H3 as of 2026-08-03 per the tracker, but the Sume route is the hosted one, so there is nothing to set up. You send a request and get a file.

Keep a note of what you asked for and what came back, including the sound. If the first clip misses, change one thing in the prompt and rerun at 480p, which costs $0.32 for 5 seconds, so a fix-and-retry loop of five runs is $1.60.

If you plan to run many clips, add an Idempotency-Key header so a retry after a timeout does not make a second job, and keep the job ids so you can fetch each result later.