MiniMax H3 vs Gemini Omni Flash for Product Videos: What We Measured

MiniMax H3 is the video model people are searching for this autumn, and the question we get is the same one we had: is it good enough for product ads, and is it cheaper than Google's? We have rendered real ads on it since October 5, 2026, side by side with Google's models, so here is what it costs, how fast it is, where it breaks, and when we still use Omni.

The Short Version

• H3 Max Turbo is less than half the price. $0.04 a second at 768p on fal, against about $0.10 a second for Gemini Omni 1.1 Flash at 720p.
• It is much faster per ad. All clips render at once: a 22-clip music ad took about 3 minutes on H3 and about 16 on Veo.
• It refused nothing. 22 of 22 clips in a full ad, and 12 of 12 in a test that included 3 shots Veo had refused.
• It doesn't talk for you. Faces turned to the camera move their mouths, and we drop H3's audio. Talking clips stay on Omni, which keeps the voice.
• It acts out every word. “No skeleton” drew a skeleton. Prompt only what the frame holds.

Price Per Second

Omni is billed in video tokens rather than seconds: a 4-second 720p clip came to 23,168 tokens in our runs, which at Google's $17.50 per million is $0.101 a second. Two traps on the H3 side. fal's pricing API reports only the 480p rate, so a cost estimate built on it is off by more than double at 768p; read the model page. And we upscale 768p to 1080p ourselves rather than paying for 1080p, which doubles the price.

Speed

A single H3 Max Turbo clip took 18 to 30 seconds in our runs; a 4-second Omni clip took 30 to 42. Per clip that is a small gap. Per ad it is large, because H3 clips can all be submitted at once and polled together, and H3 renders a clip at exactly the length the shot needs instead of a fixed 4, 6 or 8 seconds. Our first full H3 ad, a 110-second music ad of 22 shots, had its hook clip in about 30 seconds and the other 21 about 2.5 minutes later. The same ad on Veo, which we could only run one or two clips at a time, took about 16 minutes. On H3 the slowest step became our own cut, not the model.

Refusals

This is the reason we moved silent clips first. On one music-ad run, Veo refused 3 of 6 shots, flagged by its audio filter, for clips whose sound we were going to throw away. H3 Max Turbo rendered all 6, and in a wider test 12 of 12, including those 3. The full 22-shot ad that followed rendered 22 of 22. We keep a safe-prompt retry on H3 in case one is refused.

How It Looks

• Faithful to the first frame. The product, the person and the room stay what the still showed.
• Bolder camera. H3 pushes in much harder than Veo did. Good for a hook, too much for a calm product turn unless the prompt says the camera holds.
• Mouths move on to-camera shots. A person looking into the lens starts “talking” even with no dialogue. Frame silent shots with the person looking at the product, or use Omni.
• Turbo over Max for people. H3 Max cartoonified faces more than Turbo in the same test, at twice the price per second.

Prompting H3: It Acts Out Every Clause

H3 takes a prompt literally, and a negative counts as an instruction. Things that broke our clips, and the fixes:

• “X-ray only when…” and “no skeleton” both drew skeletons. Name only what the frame holds; leave out what it shouldn't.
• “Pure black” for the background turned a light-blue stage black. Say “the same colour as in the first frame” instead of naming one.
• “Warm gold light washes over” turned the whole figure gold. Mood words become paint; keep lighting words for the shot that needs them.
• “Receives a golden notification” drew a phone bubble with made-up words. Describe a glow, not a message, unless you want text on screen.

The same rule holds for Omni, less strictly. Our longer write-up on frame-bound prompts is in the scene prompt entry.

Which One to Use

How ppl.studio Uses Them

Workflows without speech, such as the drama song ad and the animated story ad, default to MiniMax H3 Max Turbo when your account has a fal key, and you can switch any of them back before the clips render. Talking workflows like the creator UGC video render on Gemini Omni 1.1 Flash on your own Gemini key, after Google shut the Veo 3.1 previews down on October 22, 2026. Images for the first frames come from Nano Banana 2.1.

Frequently Asked Questions