MiniMax H3 vs Gemini Omni Flash: Google Wins the Board by 5 Points. Then the Client Asks for a Change.

I stared at three Artificial Analysis leaderboards for ten minutes trying to work out who actually won.

Text-to-video: Gemini Omni Flash by 5 points. Image-to-video: Gemini by 2 points. Video editing: MiniMax H3 by 9 points.

Then I looked at the confidence interval column. Plus or minus 7. Plus or minus 9. Plus or minus 10.

Three boards, three gaps, every one of them inside its own error bar. In blind voting these two models are the same model.

So I closed the leaderboards and opened the dropdowns instead. That gap is not 5 points. That gap is a whole resolution tier.

Key takeaways

• All three Artificial Analysis gaps (+5, +2, +9, snapshot taken 6 August 2026) fall inside the ±7 to ±10 confidence intervals. On raw quality this is a tie, not a win.
• H3 ships 2K, up to 15 seconds, and six aspect ratios. Gemini Omni Flash's API caps at 720p, 10 seconds, and two aspect ratios. Those are enum values, not opinions.
• H3 accepts audio as an input. Gemini does not. Gemini has seed and thinking_level. H3 has neither.
• Gemini has a dedicated video-edit endpoint that mutates your source clip. H3's revision path is reference-to-video, which regenerates with your clip as a reference. They are not the same operation, and round two shows it.
• Only one of the two has downloadable weights, and it is not the Google one.

---

Round one of the MiniMax H3 vs Gemini Omni Flash test, both models animating the identical first frame, shown side by side

Round one: same first frame, same prompt, same duration. Left MiniMax H3 at 2K, right Gemini Omni Flash at 720p. Shown here as a silent GIF; both delivered files carry native audio and both are embedded as playable video further down. This is the problem. They are both good.

---

Why the MiniMax H3 vs Gemini Omni Flash Leaderboard Gap Is Being Read Wrong

Almost every MiniMax H3 vs Gemini Omni Flash post you will find quotes one leaderboard and one price. Both habits produce the wrong answer.

Here are all three Artificial Analysis boards, with the column everyone skips.

Table A: MiniMax H3 vs Gemini Omni Flash across three Artificial Analysis leaderboards (with audio), snapshot 6 August 2026

Elo figures from the Artificial Analysis Video Arena (Artificial Analysis, August 2026). Image-to-video is led by Dreamina Seedance 2.0 720p at 1,197, so neither of these two holds that crown. These are rolling crowd-vote scores and they move, which is exactly why a 5-point gap is not a verdict.

A 5-point lead with a ±9 interval means the ranking could flip next week on vote noise alone. That is not "Gemini is better at text-to-video." That is a coin toss with a scoreboard attached.

Three more things people get wrong:

The dollar-per-minute column is not what you pay. Artificial Analysis lists $6.00/min for Gemini and $7.80/min for H3. That column normalises to the cost of one minute of 1080p at default settings. Gemini Omni Flash cannot produce 1080p at all. So the number is a modelling estimate, not an invoice. Compare finished deliverables, not seconds.

One board is not the model. H3 sits second, third and first across the three boards. Gemini sits first, second and second. Quote either one alone and you can prove whatever you already believed.

"Omni" does not mean "eats anything." It is a good name and a misleading one. One of these two takes audio files as input. It is not the one with "omni" in the name.

The MiniMax H3 vs Gemini Omni Flash Spec Sheet Nobody Prints

If the Elo scores cannot separate them, the constraint tables can. I pulled both models' live input schemas rather than trusting any launch post, and the differences are not subtle.

Table B: MiniMax H3 vs Gemini Omni Flash hard constraints, read from the live API schemas on 6 August 2026

Read that table honestly and Gemini wins three rows outright. Reproducible seeds matter if you are building a pipeline that has to render the same frame twice. A real video-edit endpoint is a genuinely different tool from reference-conditioned regeneration. And thinking_level gives you a quality dial H3 simply does not expose.

H3 wins five rows, and they are the rows that decide whether you can deliver the file the client asked for.

I ran everything below on Atlas Cloud because both models sit behind the same /api/v1/model/generateVideo endpoint there, which meant one polling loop and one auth header for the whole test. Swapping models was a change to the model string. That detail is the entire reason "use both" is a realistic answer rather than a cop-out.

Table C: what each endpoint in this test costs, verified against the live price list on 6 August 2026

Neither model carries a discount this month. Gemini also exposes a cheaper developer tier ($0.112/s for text-to-video and image-to-video, $0.12/s for reference-to-video); H3 has no equivalent tier.

The open-weights line Gemini Omni Flash cannot cross. H3's weights are published on Hugging Face (MiniMax, August 2026): a 33B dense single-stream Omni Transformer, plus the Qwen3-VL-32B text encoder, a visual VAE and an audio VAE. Two caveats matter more than the download size. First, what you self-host runs at a 768-pixel short side; the Context-IR preprocessing stage and the Regenerate-2K module are not in the release, and 2K output comes from those two. Second, the community licence does not currently cover the EU, the UK, South Korea or the United States, and there is a separate application form for those territories. Read it before you build a product on it. Gemini Omni Flash has no equivalent conversation, because there is no file to download.

That open-weights position, paired with aggressive pricing, is the whole strategic story of the launch (South China Morning Post, August 2026).

Run the MiniMax H3 vs Gemini Omni Flash Revision Test Yourself

Here is the part nobody has run. Everyone benchmarks the first generation. Nobody benchmarks the second one.

Real work is not one prompt. Real work is a clip you already like, plus a client who writes back "love it, can the sign say 24H and can her apron be red." That single sentence is where these two models stop being interchangeable, and it happens to be the exact task the video editing leaderboard measures.

The scene is deliberately cruel: a rain-soaked noodle stall at night, the vendor talking straight into the lens, a hand-painted sign with legible words behind her, steam off the broth, rain on the awning. One frame that stresses lip sync, on-screen text stability, fluid motion and layered ambient audio all at once.

Both models get the identical first frame, the identical prompt, and the identical revision note. Every run below is 10 seconds, which is Gemini Omni Flash's ceiling and well inside H3's.

Step 1: Lock the first frame with GPT Image 2

Both models must start from the same image, otherwise round one measures composition luck rather than the model. Open the GPT Image 2 playground, set quality to high and ratio to 16:9, and paste this:

PlainCopy
1Photoreal night street-food stall in heavy rain, shot on a 35mm lens at f/2.0. A woman in her late thirties in an indigo canvas apron stands behind a steaming noodle counter, looking straight into the lens, mid-sentence. Behind her a hand-painted tin light-box sign glows warm amber with the words "OPEN LATE" in clean bold sans-serif capitals. Rain streaks through the sodium streetlight, steam rises off the broth pot, wet asphalt reflects red and green neon from across the street. Shallow depth of field, practical lighting only, slight lens haze, natural skin texture, no text anywhere else in the frame. 16:9.

GPT Image 2 playground on Atlas Cloud with the noodle stall prompt on the left and the generated first frame in the output panel

GPT Image 2 on Atlas Cloud: the prompt on the left, the shared first frame in OUTPUT. Cost for this step , $0.009.

The generated first frame: a woman in an indigo apron behind a steaming noodle counter at night in the rain, with a hand-painted OPEN LATE sign glowing behind her

The single input both models received. Note the two things the revision will target: the "OPEN LATE" sign and the indigo apron. Generated with openai/gpt-image-2/text-to-image.

Step 2: Round one on MiniMax H3

Go to the MiniMax H3 playground, pick the image-to-video task, upload the Step 1 frame, and set resolution to 2K, duration to 10, ratio to adaptive (the image-to-video enum offers only adaptive). Prompt:

PlainCopy
1The vendor looks into the lens and says, in a warm tired voice: "Broth's been on since four this morning. Sit down, it's still raining." She lifts the ladle as she speaks. Steam curls up across the frame. Rain keeps falling on the awning behind her. The camera holds still, no push, no pan. Ambient audio: rain on canvas, broth simmering, distant traffic, her voice close and dry.

MiniMax H3 image-to-video playground on Atlas Cloud with 2K resolution selected and the finished clip playing in the output panel

MiniMax H3 image-to-video on Atlas Cloud: 2K selected, the finished clip in OUTPUT. This capture ran at the playground's 8-second default and priced at $1.12; the 10-second version used for the comparison is embedded below and cost $1.40.

MiniMax H3, round one, 2K, 10 seconds. Sound on: the dialogue, the rain and the broth are all generated in the same pass, not layered afterwards.

Step 3: Round one on Gemini Omni Flash, same frame, same words

Open the Gemini Omni Flash playground, pick image-to-video, upload the same Step 1 frame, and paste the Step 2 prompt without changing a character. Settings: resolution 720p, duration 10, aspect_ratio 16:9, thinking_level default, seed -1.

Two things to note while you are in that form, because they are the article in miniature. The resolution dropdown contains exactly one option. The duration field stops at 10. I did not pick 720p over something better; there is nothing else to pick.

A note on what is missing here: I could not capture a run-completed playground screenshot for either Gemini step. The staging environment I take screenshots on returned 403 PERMISSION_DENIED from Google's side on all three attempts, so the only Gemini captures I have show a failed OUTPUT panel and I am not going to dress one up as a successful run. The clips below are real, generated through the production API with the settings listed above. The two H3 steps did capture cleanly and those screenshots are genuine.

Gemini Omni Flash, round one, 720p, 10 seconds, identical inputs. Play this straight after the H3 clip above and judge the audio for yourself.

Step 4: Round two, send the revision to Gemini Omni Flash

This is the round that matters. Switch to the video-edit task on the same Gemini model page, upload the clip Gemini itself produced in Step 3 as the video input, set resolution 720p and thinking_level high (a two-part edit instruction is exactly the complex case that dial is for). Paste this note exactly as a client would send it:

PlainCopy
1Keep this exact shot: same camera position, same woman, same face, same rain, same lighting, same audio. Change the hand-painted sign so it reads "OPEN 24H" instead of "OPEN LATE", in the same painted style and the same amber glow. Change her apron from indigo blue to deep crimson. Change nothing else in the frame.

Gemini Omni Flash after the revision note. Watch the sign, then the apron, then check whether anything else moved.

Step 5: Round two, send the identical revision to MiniMax H3

Here is the honest structural difference, and you should know it before you read the result. H3 has no video-edit endpoint. Its revision path is reference-to-video: you hand it the original clip as a reference and it generates a new video conditioned on that reference. It is not editing your file. It is making a new one that is supposed to look like yours.

On the MiniMax H3 page, pick the reference-to-video task, add the clip H3 produced in Step 2 to refers as a video reference, then set resolution 2K, duration 10, ratio adaptive. Paste the Step 4 prompt with no edits at all.

This step has no playground screenshot either, for a different and more boring reason: the headless browser I capture with crashed three times on the reference uploader while loading a video reference. The run itself went through the API without complaint.

MiniMax H3 after the identical revision note, via reference-to-video at 2K.

Side-by-side of round two: the Gemini Omni Flash video-edit result next to the MiniMax H3 reference-to-video result, both from the same revision note

Round two side by side, silent GIF. Left Gemini Omni Flash via video-edit, right MiniMax H3 via reference-to-video. Both had the same sentence to work from.

What actually happened in round two. I went in expecting H3 to lose this one. Reference-conditioned regeneration is a blunter instrument than a real edit endpoint, and "regenerate the whole clip and hope it lands in the same place" is exactly how you get a different face, a drifting camera and a client asking what happened to the shot.

That is not what came back. Both models did the job, and both did it cleanly.

Gemini's video-edit behaved exactly as advertised. The sign reads OPEN 24H in the same hand-painted lettering, with the same amber glow and the same weathering on the tin. The apron is deep crimson. Everything else is untouched: same face, same expression, same ladle angle, same steam shape, same rain, same tail lights in the background. It reads as the original file with two pixels-deep corrections, because that is essentially what it is.

H3's reference-to-video produced the same two changes and held the shot far better than the architecture suggests it should. Sign changed, apron changed, and across all 10 seconds the framing, the pour of broth off the ladle, the steam plume and her face stayed on model against the round-one clip. If there is drift here, I could not find it by stepping through frames.

So round two is a third statistical tie, and I am not going to pretend otherwise to make a tidier story. What separates the two outputs is not the edit quality. It is that H3 handed back 2560x1440 with a 32 kHz stereo track and Gemini handed back 1280x720. Same instruction, same success, different deliverable.

One honest caveat on method: this is a single revision on a single shot, not a controlled study. A two-part change to a sign and a garment is a fairly friendly edit. Harder cases (removing an object the camera moves past, changing something the subject occludes, four rounds of notes stacked on each other) are where a dedicated edit endpoint should pull ahead, and where regeneration should start to wobble. Run yours before you decide.

Three More MiniMax H3 vs Gemini Omni Flash Splits Worth Testing

Round two is the split that changes a delivery. Three more are worth twenty minutes of your own credit.

Feed it an audio file. H3's reference-to-video accepts up to three audio clips of 2 to 15 seconds each, as long as at least one image or video reference comes with them. That means you can hand it a real voice recording and have the character perform it. Gemini Omni Flash has no audio input field on any of its four endpoints, so this comparison cannot be run at all. That is not a scoring loss for Google; it is a capability that is simply absent.

Ask for 21:9. H3's reference-to-video ratio enum contains 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Gemini's aspect_ratio enum contains 16:9 and 9:16. If your deliverable is a widescreen title sequence or a 1:1 social cut, one of these two models is not in the conversation.

Lock a seed and re-run. This one goes the other way. Gemini exposes seed; H3 does not expose it on any endpoint. If you are building a pipeline where the same request has to return the same frames on Tuesday that it returned on Monday, Gemini gives you a handle and H3 gives you nothing. For regression testing, A/B copy variants, or any automated render farm, that is a real advantage and it belongs on the Google side of the ledger.

What a MiniMax H3 vs Gemini Omni Flash Test Actually Costs

The whole experiment above, four video generations and one image, came to about $5.51. One first frame at $0.009, two 10-second round-one clips at $1.40 and $1.30, two 10-second round-two clips at $1.40 each.

Per second, Gemini is cheaper: $0.125 to $0.14 against a flat $0.14 for H3, so roughly 7 to 11 percent less depending on the endpoint. That number is true and it is also nearly useless on its own.

Here is why. Suppose the deliverable is a 15-second 21:9 cinematic opener at 2K. H3 renders it as one clip. Gemini cannot render it at all: not the resolution, not the duration, not the aspect ratio. You would be stitching a 10-second and a 5-second 16:9 clip, cropping to fake the widescreen, and shipping 720p. The per-second price of a thing you cannot deliver is not a saving.

Flip it. Suppose the deliverable is a 16:9 social cut that will go through four rounds of client notes and has to come back byte-identical when the legal team asks to re-render it in three weeks. Gemini's video-edit plus seed is built for exactly that loop, and it costs less per second while doing it.

Table D: which MiniMax H3 vs Gemini Omni Flash job goes where

The conclusion of this article is that table, not a score. If a benchmark cannot separate two models by more than its own error bar, the benchmark has told you everything it knows, and the spec sheet takes over from there.

For a wider view of where H3 lands against the rest of the field, the MiniMax H3 alternatives breakdown covers the models that beat it on the boards it does not lead.

Frequently Asked Questions

Is MiniMax H3 better than Gemini Omni Flash?

Blind voting cannot separate them. Across three Artificial Analysis leaderboards the gaps are 5, 2 and 9 Elo points, and every one sits inside a ±7 to ±10 confidence interval. Pick on specifications, not on rank. The resolution ceiling, duration cap and aspect ratio list will change your deliverable; 5 Elo points will not.

Which is cheaper, MiniMax H3 or Gemini Omni Flash?

Per second, Gemini. On Atlas Cloud it runs $0.125 to $0.14 per second against H3's flat $0.14, with a developer tier at $0.112 per second that H3 has no equivalent of. But Gemini is capped at 720p, 10 seconds and two aspect ratios, so for many deliverables you are not comparing the same product. Price the finished cut, not the second.

Can Gemini Omni Flash generate 1080p, 2K, or 15-second video?

No. Its API resolution parameter has one legal value, 720p, and duration accepts 3 to 10 seconds. MiniMax H3 offers 768P or 2K and 4 to 15 seconds. These are enum constraints read from the live schema on 6 August 2026, not editorial opinion, and Google may lift them later.

Can I self-host Gemini Omni Flash the way I can self-host MiniMax H3?

No. H3's weights are on Hugging Face and run locally at a 768-pixel short side. The Context-IR stage and the Regenerate-2K module that produce 2K output are not in the release and stay API-side. Also check the licence: it does not currently cover the EU, UK, South Korea or the US without a separate application.

Do I have to pick just one?

No, and the split-by-job table above is the better answer. Both models sit behind the same generateVideo endpoint on Atlas Cloud, so running both means one polling loop and a different model string, not two integrations. Routing by deliverable beats betting on a leaderboard that moves every week.