Picking a minimax/h3 tier: a decision flow from 480p to 2k

Updated 2026-10-02

The tier explainer shows how per-second pricing works across tiers. This page is the other half: given a project, which tier do you request? The model page for minimax/h3 lists tiers such as 480p, 540p, 768p, 1080p, 1440p, 2k and 2160p, plus variants with an -esr suffix (for example 1080p-esr); other pages may show -sr variants. The pages do not define those suffixes in a way this site can quote, so read the note on the model page and treat suffixed rows as separate, differently priced tiers rather than guessing what they do.

Step 1: decide the delivery target first

Work backwards from where the clip will be seen, not forwards from what the model can produce. A clip played in a phone feed or an embedded player does not benefit from a tier the viewer's screen cannot show. A clip that will be cropped, reframed or scaled up in editing needs more pixels than its final frame.

These are starting hypotheses, not claims about quality. Whether the next tier up is visibly better for your content is something you verify with the test below.

Step 2: the decision flow

1. Is this a draft? If yes, use the cheapest tier available and a short duration_secs. Stop.
2. Will anyone but your team see it? If no, stay at the draft tier.
3. Is the final display larger than 1080 pixels on its long side? If no, test 1080p as your ceiling.
4. Will it be cropped or stabilised in post? If yes, add one tier of headroom.
5. Is the deliverable contractual (broadcast spec, ad platform requirement)? If yes, use the tier that satisfies the spec and verify the returned file against it.
6. Otherwise, run the paired test below and pick the lowest tier your reviewer cannot distinguish from the next one up.

Step 3: the paired test

Choose three representative prompts. Render each at two adjacent tiers with identical prompt, duration and aspect ratio. Show the pairs, unlabelled and in random order, to whoever approves output, and ask which they would ship. If they pick the lower tier or cannot tell, you have your answer. This costs a few clips and avoids paying a per-second premium on every future job.

import os, random, time, requests

API = "https://videorouter.sh/api/v1"
H = {"Authorization": f"Bearer {os.environ['VIDEOROUTER_KEY']}"}
TIERS = ["768p", "1080p"]          # adjacent tiers to compare
PROMPTS = ["slow dolly through a market at dusk", "..."]

def run(prompt, tier):
    j = requests.post(f"{API}/videos", headers=H, json={
        "model": "minimax/h3", "prompt": prompt,
        "resolution": tier, "duration_secs": 5}).json()
    while j.get("status") not in ("completed", "failed"):
        time.sleep(5)
        j = requests.get(f"{API}/videos/{j['id']}", headers=H).json()
    return j

results = [(p, t, run(p, t)) for p in PROMPTS for t in TIERS]
random.shuffle(results)             # blind order for the reviewer

Step 4: confirm you received the tier you asked for

Resolution is a hint the model may not honour: an unrecognised or unsupported combination is ignored and the default is used, and the request still returns 200. So measure:

ffprobe -v error -select_streams v:0 -show_entries stream=width,height -of csv=p=0 out.mp4

If the dimensions do not match your request, either the host does not support that tier or the aspect ratio changed the dimensions. Pixel dimensions at a given tier name also depend on aspect ratio, so compare against the model page rather than assuming 16:9.

Cost and routing considerations

Higher tiers cost more per second, and the spread between the cheapest and priciest host differs by tier. The live table shows the range for minimax/h3:

Per second, before VideoRouter's 2% platform fee. For tiered models each row compares the resolution tier with the widest host-to-host gap. Built 2026-10-02 from the live catalog.

• Unpinned requests go to the cheapest healthy host. Not every host offers every tier, so a tier that only one host serves effectively ties you to that host's price.
• A suffix such as minimax/h3/<host> is a soft preference and can still fall back; a hard pin needs provider.only with allow_fallbacks: false.
• You are billed once at creation from the requested duration at the chosen tier, polling is free, and jobs that fail upstream are not billed. The platform adds a 2% fee on video.

Common mistakes

• Developing at the delivery tier, which multiplies the cost of every prompt iteration.
• Assuming the highest tier is always safest. It costs more and, depending on your delivery size, may add nothing visible.
• Comparing tiers across different hosts, so a difference you see may come from the host rather than the tier. Pin one host for the comparison.
• Skipping the dimension check, then discovering the default tier was used for an entire batch.

Next, read the prompt guide to get more from each render, browse the model pages for the tier rows on each host, or create a key and run the paired test.

Frequently asked questions

Keep reading

• MiniMax H3 Providers Compared — How to Pick a Host
• MiniMax H3 Image-to-Video API: start_image_url Guide
• MiniMax H3 Python Tutorial: Create, Poll, Download, Pin
• MiniMax H3 Resolution Tiers: How Per-Tier Pricing Works