Alternatives

MiniMax H3 vs Seedance 2: three versions are live, and only one of them is priced

The Seedance people put against H3 is 2.5 — which has no published rate above 720p and no score on any public blind-test board. That is a planning problem, not a quality claim.

Aug 8, 20268 min readEditorial deskEditorial desk
MiniMax H3 vs Seedance 2: three versions are live, and only one of them is priced

This is three comparisons wearing one name, because ByteDance has three live versions and MiniMax has one.

On list prices you can check today, H3 runs $0.08 per output second at 768P and $0.13 at 2K. Seedance bills by token, so its cost climbs with pixel count: $0.154 per second at 720p on 2.0, $0.231 on 2.5. Above 720p, 2.5 has no published rate at all.

Which Seedance are you comparing?

Start here, because H3 versus Seedance 2.5 is not the same comparison as H3 versus Seedance 2.0.

Seedance 2.0Seedance 2.1Seedance 2.5
ShippedFebruary 202631 July 2026
ResolutionFour tiers, up to 4KBarely documented480p and 720p published
Length4–30 s in one pass
ReferencesUp to 50 multimodal
Published rate above 720pYesNone
Blind-test scoreYesNone

Two rows are worth reading twice. BytePlus publishes 2.5 rates for 480p and 720p only, and its own product page lists those two resolutions. And no public blind-test board has scored 2.5 — not Artificial Analysis, not arena.ai — while ByteDance shipped it without a benchmark chart.

One more trap sits underneath all of it: the China region runs a separate contract in yuan under a different model tag, so converting at the exchange rate does not give you the international number.

None of that makes 2.5 worse. It makes it unbudgetable and unranked, which is a different problem and yours to plan around.

Per output second against tokens per pixel

The two bills are shaped differently, which is why "cheaper" changes answer by resolution.

H3 charges a flat rate per output second, two tiers, and charges reference video by its own length at the same rate. Seedance multiplies tokens: input seconds plus output seconds, times width, times height, times frame rate, divided by 1024, then times the token rate. Pixels sit inside that formula, so cost climbs roughly with area, and a minimum-token floor can override the whole calculation on short clips.

Run it and the crossover sits between 480p and 720p. At 480p, Seedance 2.0 is about ten per cent cheaper than H3's 768P, and its Mini tier is less than half. At 720p and above, H3 wins every tier: $0.08 against $0.154 on 2.0, and against $0.231 on 2.5, for a short side of 768 pixels against 720.

At the same delivery size the price is not close. It is also not a quality claim.

Whose price is that, exactly

Here is the part that quietly ruins budgets. Ask four sources what H3 costs at 768P and you get $0.05, $0.08, $0.14 and $0.16 — a 3.2× spread on one model, because three of those are resellers quoting their own markup and one is MiniMax's list rate.

The same happens on the other side. BytePlus lists 720p on 2.0 at $0.154 per second; resellers of the same endpoint quote $0.134, $0.20 and $0.3034. Consumer apps are a third shape: they quote credits, which do not convert cleanly into seconds.

One of those resellers is cheaper than the vendor. That is allowed, and it still is not the price of H3. Before you quote a number from anywhere, including here, check which row it came from.

On the blind-test boards

Artificial Analysis runs blind human-preference voting and publishes three video boards. Read all three, because the two models split them.

BoardMiniMax H3Seedance 2.0Verdict
Video editing with audio1,125 (11,445 votes)1,033 (21,335 votes)H3, clearly
Text to video1,2381,222Inside the interval — a tie
Image to video1,1931,198Inside the interval — a tie

Ninety-two points on the editing board is far outside both confidence intervals, so that one is a real result. The other two gaps are smaller than the intervals around them, which means neither is a win.

Seedance 2.5 is on none of these boards. Any Elo you see attached to it was invented somewhere.

Who has actually run both

Four public paired runs exist, and we are listing all of them, including the one that goes against us.

On 11 August a reseller that sells both models ran six generations, five seconds each, same prompt. Seedance 2.0 came back with cleaner dialogue staging, better product identity and more convincing complex motion. It was also faster, a median of 137 seconds against H3's 179. H3 cost 75 per cent less. Both delivered three usable clips out of three.

A second run found something more useful than either verdict: fed one composite character sheet and one lighting board, both models held the same face. Character consistency turned out to be a packaging problem, not a generation-number problem.

Length, references and how each one edits

Three real capability gaps, and one that is smaller than it looks.

Length is the clearest. H3 gives you 4 to 15 seconds in one pass; Seedance 2.5 gives you 4 to 30, with an experimental mode reaching further. If your cut needs one unbroken half-minute, that is decided.

Editing granularity is next. H3 takes whole-clip instruction edits: swap the background, rewrite the line, keep the rest. Seedance 2.5 redraws a region and leaves performance, lighting and camera alone. Both write sound in the same pass as the picture.

Then references, where the headline gap is fifty files against twelve. The paired run above suggests the count is rarely what limits you: one well-built composite beats six loose crops on either model.

When Seedance beats H3, and when not to use this site

Six situations where we would send you the other way.

Need 4K, or one continuous shot past fifteen seconds, or a region-level fix inside an approved cut: Seedance, and check your platform's access first. Working inside CapCut or Jimeng already: Seedance, because the handoff is native. Drafting hundreds of 480p tests: Seedance's Mini tier, at less than half H3's rate. Budgeting a 1080p or 4K deliverable on 2.5: nobody can price that yet, so plan on 2.0.

Two more that are about us rather than the models. If you need commercial rights to what you generate, do not generate it here, because every plan on this site, free included, is personal and evaluation use. And if self-hosting matters, H3's weights are excluded from four markets, which is its own subject.

The questions people ask first

Which should I pick? Depends on the board and the budget. H3 wins video editing clearly, the other two boards are ties, and H3 is cheaper at every tier from 720p up.

Which Seedance version should I compare against? The one your platform can actually call. 2.0 if you need a price above 720p or a blind-test score; 2.5 if you need length.

Is MiniMax H3 cheaper than Seedance? At 720p and above, yes, by roughly half. At 480p, no.

Why do I keep seeing different prices for H3? Because most of them are resellers quoting their own rate. MiniMax's list price is $0.08 per second at 768P and $0.13 at 2K.

How long can each one go? H3 does 4 to 15 seconds per pass. Seedance 2.5 does 4 to 30.

Does Seedance 2.5 really do 4K? ByteDance's launch materials say so. Its product page lists 480P and 720P, and no rate is published above 720p.

What is Seedance 2.5's arena score? There isn't one. No public board has rated it.

How does either compare with Veo 3.1? Different question, different page: how H3 compares with Veo covers that one.

What do the two have in common? Both write their own audio in the same pass as the picture, both bill per unit of output, and both take mixed references.


Every price on this page is a vendor's list rate. Ours is on the pricing page. Price it out, then generate one clip and judge the picture.

Editorial desk

Written by

Editorial desk

minimax-h3ai.video

All articles