Two frontier video models, one shared prompt
Category
Growth
Published Date
ElarisLabs Marketing
Summary
Seedance 2.5 and MiniMax H3 launched on the same day. We read both spec sheets, listed what each actually ships, and ran one prompt designed to break things through both.
Share this article
Seedance 2.5 and MiniMax H3 launched on the same day at the end of July. Both generate audio in the same pass as picture, both take large reference packages, and both arrived with a wall of demo reels that prove very little. So we did the boring thing: read both spec sheets properly, listed everything each one actually ships, then wrote a prompt designed to break things and ran it through both.
Same launch day. Different companies, different bets.
Before any judgment about quality: what each one is, who builds it, when it shipped, everything it publicly claims to do, and where you can actually get at it.
Seedance 2.5 (ByteDance Seed, launched 31 July 2026)
Built by ByteDance Seed, ByteDance's research division. Previewed at the Volcano Engine FORCE conference in late June and shipped at the end of July. ByteDance skipped 2.1 through 2.4 entirely, framing this as a generational jump rather than a point release. The pitch is long-form: a whole scene, with its own internal shot changes, out of one generation.
Duration: up to 30 seconds in a single pass, with multi-round extension beyond that. No stitching and no extension pass needed to reach half a minute.
Multi-shot: organises several logically connected shots inside one generation, with improved transitions and scene changes, so a story unfolds without a cut you have to assemble.
References: up to 30 images, 10 video clips and 10 audio clips in a single pass — 50 reference inputs total.
Reference types: character, motion, clay-render and creative references, plus R2V support with green-screen and white-model references for blocking.
Editing: timestamp-level control for targeted edits to audio and video without regenerating the whole shot.
Audio: native joint audio-video generation, inherited from the Seedance 2.0 architecture.
Camera: professional camera movement and performance blocking, including 3D camera blockout controls.
Weights: closed. Platform and API access only.
Resolution: announced with native 4K. Verify inside your own workspace before planning around it — see the note below.
MiniMax H3 / Hailuo 3.0 (MiniMax, launched 31 July 2026)
Built by MiniMax, the lab behind Hailuo 02 and Hailuo 2.3. Officially named MiniMax H3; Hailuo 3, Hailuo 3.0 and Hailuo 03 all point at the same model. The pitch is that it stops behaving like a video model — one transformer reading text, images, video and audio as a single context.
Duration: 4 to 15 seconds per generation at 24fps, with platform tooling to extend further.
Resolution: 768p and 2K, native — not a separate upscale pass. The 1440p tier circulating in the community is not an official option.
Audio: 32 kHz native stereo generated in the same pass as picture: dialogue, effects and room tone, timed to on-screen action. No separate audio stage.
References: up to 9 images, 3 video clips and 3 audio clips — 12 files total — usable for identity, motion, framing, atmosphere or voice without forcing them as keyframes.
Modes: text-to-video, first- and last-frame image-to-video, and reference-to-video with mixed inputs.
Editing: instruction-based. Request a change in a sentence instead of rerolling the shot. Also does video-to-video motion transfer and voice transfer.
Prompt limit: 7,000 characters.
Weights: open. A 33B dense single-stream Omni-Transformer, published on Hugging Face in early August 2026 under the MiniMax H3 Community License.
Benchmarks: ranked #1 in Video Editing, #2 in Text-to-Video and #3 in Image-to-Video on the Artificial Analysis leaderboard at launch.
Read both spec sheets with a hand on the brake
Both launches are recent enough that the public record still contradicts itself. Seedance 2.5's announced native 4K has been reported as not making it into the shipped model, with some routes topping out well below it. MiniMax marketed H3 as open-weight before any weights existed, so comparison pieces published in that window state flatly that neither model is open — those sections are now stale.
Access is layered too. On Seedance, "available" means one thing inside Dreamina or Jimeng and something else on BytePlus ModelArk and Volcano Engine, where the developer contract landed after the consumer product. Verify the model label inside your own signed-in workspace before you plan a pipeline around a number you read in a blog post. Including this one.
The scene model and the shot model
Read down the table below and the strategic difference is obvious. Seedance is betting the unit of work is a scene. MiniMax is betting it is a shot. That decides where each one sits in a pipeline more than any quality score will.
Published specifications, August 2026 (Seedance 2.5 / MiniMax H3)
Company: ByteDance Seed / MiniMax
Launched: 31 July 2026 / 31 July 2026
Predecessor: Seedance 2.0 / Hailuo 2.3
Max single generation: 30 seconds / 15 seconds
Extension: multi-round extension / platform extend tooling
Frame rate: not officially published / 24 fps
Native audio: yes / yes — 32 kHz stereo
Image references: up to 30 / up to 9
Video references: up to 10 / up to 3
Audio references: up to 10 / up to 3
Multi-shot in one take: yes, with scene changes / yes, within 15s
Editing model: timestamp-level / instruction-based
Prompt limit: not published / 7,000 characters
Weights: closed / open — 33B, community license
Billing unit: tokens, via provider / per generated second
50 Seedance reference inputs, max
12 H3 reference files, max
33B H3 open-weight parameters
0 days between the two launches
You cannot divide one of these prices by the other
This is the part most comparisons get wrong. The two models do not bill in the same unit, so any figure produced by comparing them directly is meaningless.
MiniMax publishes a straightforward per-generated-second rate for H3 — roughly $0.13 per second at 2K and $0.09 per second at 768p, with extra charges for excess image references and for input video duration. A 15-second reference clip is billed as fifteen seconds of input on top of your output.
Seedance 2.5 bills through token-based provider rates that change depending on whether video input is present, and the numbers differ across Volcano Engine, BytePlus ModelArk and aggregators. Video-input duration feeds into token consumption. There is no single published per-second figure to hold against H3's.
The metric that actually decides anything:
Under that framing the answer flips depending on the work. A short, bounded 2K shot is cheap to review and cheap to throw away, so H3 wins on iteration economics. One accepted 30-second scene can remove four generations, three stitch points and a continuity pass, so Seedance wins when the retry rate is low and the scene is long.
If you are testing prompts, H3 is cheaper to fail at. If you are delivering scenes, Seedance is cheaper to succeed at.
Method: we tested at fifteen seconds. Neither model stops there.
Seedance 2.5 will run to thirty seconds in one pass. H3 will run to fifteen. Testing at fifteen is not a statement about what either model can do — it is the only window where the two are doing the same job, and therefore the only window where a result means anything.
Running Seedance at thirty and declaring it the winner measures a spec sheet, not a model. Long-form generation, multi-shot continuity across half a minute, and Seedance's timestamp-level editing all deserve their own test, and they will get one. This one is about what happens inside the overlap.
Everything else is held constant: 16:9, the same two reference images attached to both runs, one generation each, no rerolls. Both outputs at each platform's highest available native resolution, downscaled to a shared 1080p canvas only at the comparison stage, with both original files retained.
Why this prompt
Detail alone stopped discriminating between frontier models a while ago — both of these will render a beautiful static portrait. To learn anything you have to attack the parts that still break, and you have to be able to see the break while scrolling, not in a lab. Every beat below is independently gradeable.
00:00 – 00:03 · Bare face on the ledge
Close on his unmasked face, rain on the skin, breath visible, neon across one cheek. The camera eases back to reveal the helmet held in both gloved hands, weight pulling the wrists down.
Watch for: whether the face matches the portrait reference at all, and whether the helmet has believable mass in the hands or floats like a decal.
00:03 – 00:05 · The helmet goes on (key beat)
He raises it, turns it once so the visor throws the skyline back in curved miniature, then lowers it over his head — rim past the ears, hair pressing flat, shell seating at the collar, a roll of the neck to settle it.
Watch for: hands manipulating a rigid object with occlusion. Fingers should grip a hard shell, not squash it. No clipping through the skull, no skin visible once seated.
00:05 – 00:06.5 · The hand sign, in macro (key beat)
Camera drops to the right hand. Index and little fingers snap straight up and spread; middle, ring and thumb fold flat to the palm. Glove leather creases where the knuckles fold. The wrist cocks back and holds.
Watch for: a specific held finger configuration at close range. The hardest thing in the shot. Count the digits, check the joint bends, check whether the pose survives the beat or quietly morphs.
00:06.5 – 00:08 · Twin strands, braiding (key beat)
Macro on the two raised fingertips, rain beading on the glove. The wrist snaps and two separate web strands fire in parallel, braiding into one taut line a short distance out. Recoil travels back through fingers, wrist, forearm.
Watch for: two strands actually converging. Models cheat two ways — collapsing to one strand immediately, or letting the pair drift and never merge. The line should have thickness and sag, not glow.
00:08 – 00:11 · Swing, car roof, wall run
Wide on a brownstone street with a neon MARA'S DELI sign. He swings low, boots skimming a parked car roof with real compression, redirects off it, then lands feet-first on a wall and runs four steps along it before pushing off and firing again.
Watch for: the sign staying spelled correctly, and whether the car roof and the brick push back. Floating contact is the most common failure at this speed.
00:11 – 00:13.5 · Street, taxi, launch
Low over a street of yellow taxis, headlights smearing on wet asphalt. The web releases, boots hit pavement for two hard strides, a horn sounds, he plants a boot on a taxi and pushes off the roof to rocket up. Then a one-second hold at the apex.
Watch for: reflections on wet asphalt moving coherently with the camera rather than sitting painted in place. Also whether the splash lands on the same frame as the footfall.
00:13.5 – 00:15 · Landing, unmask, the smile (key beat)
He lands in a low crouch, one gloved hand on wet concrete, knees absorbing. He rises, grips the helmet with both hands and lifts it clear. Hair pushes free, damp. He looks out over the city and breaks into a real smile.
Watch for: the same face. Produced at second one, hidden for eleven seconds under a rigid shell, reproduced at second fourteen. Also the landing — models float the touchdown instead of absorbing it.
Which to use: they are not competing for the same job
Neither model has been independently benchmarked long enough for a verdict to survive the next release cycle. What follows is a fit decision, not a ranking.
Reach for MiniMax H3 when
The deliverable fits inside fifteen seconds
You want 2K without an upscale pass
On-screen type is part of the brief
You are iterating heavily and need cheap failures
You want instruction-based edits instead of re-rolls
Compliance needs weights outside a vendor API
Reach for Seedance 2.5 when
The scene needs more than fifteen seconds to breathe
You want multiple shots inside one generation
You are stacking large reference packages for character consistency
Motion, clay-render or white-model references matter to your blocking
Timestamp-level editing removes a real revision bottleneck
You already work inside the Dreamina, Jimeng or CapCut stack
And the answer most production teams will actually land on is both. Seedance for the establishing scene, H3 for the inserts and coverage. They are not competitors in a pipeline. They are different tools that happened to launch on the same day.
Run it yourself. Tell us we're wrong.
ElarisLabs is the AI-native creative operating system for brands. We publish the tests we run internally — including the ones where the model we expected to win didn't — because a comparison you can reproduce is worth more than a reel you can only admire.
