MiniMax H3 vs Seedance: The Per-Second Math Nobody Prints

Every comparison of MiniMax H3 vs Seedance currently on the first page settles on the same verdict: H3 is the value pick, Seedance is the quality pick. That verdict is fine. What none of them print is the number it rests on, and the number turns out to depend on something nobody mentions: the shape of your frame.

Here is what a take actually costs on each, taken from the model catalogue our own generator reads at runtime.

The short answer

• Shooting longer than about eight seconds → MiniMax H3. The per-second bill stops scaling against you.
• Shooting short and wide, or short and square → Seedance 2.0. Under five seconds the gap is small enough to ignore, and a square frame is cheaper than anything H3 offers.
• Need sound in the same generation → Seedance 2.0. H3 comes back silent through our endpoint.
• Need 4K → Seedance 2.0, at 778 credits for five seconds and an active subscription.
• Need a single take past fifteen seconds → Seedance 2.5, the only one that goes to thirty.
• Need it back faster than playback → H3 Max, a different model from H3 despite the name.

Side by side

That table is the whole of MiniMax H3 vs Seedance in nine rows, costed against our production catalogue on 9 September 2026. Credit figures come from the same function the generator calls before it submits a job, so they are what you would actually be charged, not an estimate.

Two billing models, and only one of them cares about your frame

This is the mechanism the whole comparison turns on, and it is why a single "price per second" for Seedance cannot exist.

MiniMax bills per second of output. Five seconds costs the same 130 credits whether you render 21:9, 16:9, 1:1 or 9:16. The frame is free.

Seedance bills per pixel. The catalogue computes tokens as width × height × 24 / 1024 and charges a rate per thousand. So the same five seconds at the same 720p label:

A 21:9 Seedance take costs 2.3 times a 1:1 one at identical resolution and duration. Against H3 that flips the verdict twice inside one table: cinemascope is 53% more expensive on Seedance, and a square social crop is 34% cheaper.

Any article that gives you one Seedance price per second has quietly picked an aspect ratio for you and not said which.

What a take actually costs

The interesting comparison is not five seconds. It is what happens as the take gets longer, because the two curves have different slopes.

At five seconds and 720p, Seedance costs 17% more than H3 and gives you a soundtrack for it. That is close enough that quality preference should decide, not price.

Compare at matched quality instead and it stops being close. H3's 2K sits between Seedance's 720p and 1080p tiers, nearer to 1080p. A fifteen-second 1080p Seedance take is 1,021 credits; the same fifteen seconds at 2K on H3 is 390. That is 62% less for more pixels. Nothing about the per-second headline tells you this, because the gap is created by the tier you have to climb to on Seedance to reach comparable sharpness.

Both scale linearly with duration, so there is no volume break waiting further out in either direction. Our pricing page has the credit-to-currency side.

The audio difference is real, and usually stated wrong

Search results will tell you MiniMax H3 generates native stereo in a single pass. That is true of the model. It is not true of the clip you get from us.

Our catalogue marks minimax-h3 as supportsAudio: false. The render arrives silent. Meanwhile H3 Max is marked audioAlwaysOn — every clip carries sound and there is no switch to turn it off.

Same family, opposite behaviour, and both descriptions are accurate at the layer they describe. Capability belongs to the model; behaviour belongs to the endpoint. Most of the "the specs I read elsewhere do not match what I got" confusion in this space comes from collapsing those two, which is why we keep a separate page sorting out the H3 name family.

Seedance is cleaner here: audio is supported on both versions and ships off by default, so you opt in per generation at no extra cost.

MiniMax H3 vs Seedance in reference mode

Reference mode reshuffles the arithmetic, in opposite directions on each side.

H3 charges for reference images. Seedance does not. The first five are included; each one after that adds to the bill. Seedance counts references as input and charges nothing for them.

Seedance gets cheaper in reference mode. Our 2.0 reference endpoint runs the fast tier, about 20% below the standard token rate, at the cost of a 720p ceiling. Text-to-video and image-to-video keep the higher tier and the higher ceiling.

So the ordering flips depending on how many stills you feed it. At five references H3 and Seedance 2.0 are within nine credits of each other. At nine, H3 costs 34% more than the Seedance take it was cheaper than a moment ago, and you are still capped at 2K against 720p.

If you work with reference packs, this is the line that decides your bill, and it is missing from every comparison currently ranking. Both reference workflows have their own page: Seedance and the H3 family.

The fifty-reference figure is wrong, and it is spreading

AtlasCloud's comparison, currently on the first page for this query, credits Seedance 2.5 with "up to 50 multimodal references." Several other results repeat the number.

Fifty is the combined ceiling across images, video and audio. The image slot alone is 30. Our reference endpoint sends images and nothing else, so 30 is both the catalogue limit and the real one, and you address them individually in the prompt as @Image1, @Image2. If you plan a shot around fifty stills you will hit a wall twenty short of it. The reference-to-video page has the working limits.

Worth saying plainly, because the same article is otherwise the most honest one on the page: it discloses in its own text that Seedance 2.5 was not open on its platform when the test ran, so the Seedance side of that comparison is Seedance 2.0. A test nobody else disclosed, disclosed. The reference count is a sourcing error, not a credibility one.

For contrast, ByteDance's own launch post states the 2.0 input budget as nine images, three video clips and three audio clips, which is the shape the fifty-figure got flattened out of.

What the local-versus-cloud gap actually is

Coverage of H3 leans on it being open-weight, and it is. The part that gets dropped: official 2K runs through a regenerate module that has not been open-sourced. Pull the weights and your ceiling sits well below the 2K number in the spec sheet.

So "run it yourself for free" and "2K output" are both true statements about MiniMax H3 that are not true at the same time. Running locally is genuinely free and we are not; that trade is real, and it is the honest reason to skip a hosted endpoint. Just do not budget a local pipeline around a resolution the open weights do not reach.

Both fal model pages carry the current hosted rates: H3 and H3 Max, the latter running a launch discount that expires 14 September 2026.

The cheap line neither comparison mentions

Both models above are the wrong tool for drafting, and the catalogue has a third answer that no MiniMax H3 vs Seedance article includes, because it is not really either of them.

A five-second 768p take costs 20 credits on H3 Max Turbo and 40 on H3 Max. Against 130 for H3 at 2K and 152 for Seedance at 720p, that is between six and eight times cheaper than the models people actually compare.

The catch is real: 768p, no reference mode, and audio you cannot switch off. It is not a finishing tier. But for testing whether a prompt holds up before you spend 1,021 credits on a fifteen-second 1080p take, the arithmetic is hard to argue with. Both run in Live Mode, where the render comes back faster than the clip plays.

Where they are the same

Enough of the MiniMax H3 vs Seedance framing is marketing that the overlaps are worth stating.

Both cap a single take at fifteen seconds. Both accept nine reference images. Both take the same six aspect ratios, with H3 adding an auto option in reference mode. Both bill per generation with no seat cost. Neither is reachable through an API on our platform, because our generate endpoint only accepts a signed-in session.

The differences that matter are the three above: the billing shape, the audio behaviour, and the resolution ceiling. Everything else is close enough that a preference test beats a spec sheet.

Which one to pick

• Take runs past eight seconds → MiniMax H3. The per-second bill is what saves you, and the 2K comes along for free.
• Take is under five seconds and needs sound → Seedance 2.0 at 720p, 152 credits, audio switched on.
• Square or vertical social crop → Seedance 2.0. Pixel billing works in your favour on narrow frames: 86 credits at 1:1.
• Cinemascope → MiniMax H3. Pixel billing works against you at 21:9, where Seedance charges 199 against H3's flat 130.
• 4K final → Seedance 2.0, 778 credits for five seconds, subscription required.
• One take past fifteen seconds → Seedance 2.5, the only one that reaches thirty.

Most MiniMax H3 vs Seedance write-ups stop at a verdict. The verdict is the easy part; the frame shape and the reference count are what actually move your bill. If you are choosing between the two Seedance versions rather than across families, the 2.0 versus H3 comparison covers that side, and it is a genuinely different trade: 2.5 is not an upgrade, it is longer takes and more references bought with resolution and price.