Table of Content
| Model | Max native clip | Longer than that? |
|---|---|---|
| SEEDANCE 2.5 | 0:30 one pass | Extend a finished clip twice |
| MINIMAX H3 | 0:15 | Stitch clips in an editor |
| WAN 3.0 | 0:30 one shot | Separate extension tool |
The first week of August turned into an unplanned lab experiment for me. Seedance 2.5 and MiniMax H3 launched on the same day, July 31, and Alibaba opened the Wan 3.0 public beta six days later. Three major video models inside one week is not normal, so I cleared my calendar, loaded credits on Dreamina, wired up the MiniMax API, and borrowed a colleague's Alibaba Cloud access for Wan's invitation-gated beta.
This is not a spec sheet rewrite. I pushed the same four briefs through each model: a 30-second coffee brand spot, a vertical product clip with on-screen text, a two-character dialogue scene, and a round of edits on footage I already liked. I logged every cost, queue time, and first-take success rate.
The short version: these three are not really fighting for the same job. Seedance 2.5 is built for long, directed scenes. H3 is built for fast, affordable short-form with sound. Wan 3.0 is an ambitious beta with one trick nobody else has. The long version, with numbers, is below.
Pick by job, not by hype
Best for storytelling Seedance 2.5 One 30-second take, up to 50 reference files, and lip sync across 10+ languages. The model I trust when a clip needs a beginning, a middle, and an ending. | Best value and access MiniMax H3 Native 2K with stereo sound at $0.13 per second, open weights, and the strongest post-render editing I tested. My daily driver right now. | Best to watch, not buy Wan 3.0 A 30-second single take plus a document-to-video mode with no rival, locked in an invitation-gated beta. Promising and unfinished in equal measure. |
Which of these actually exists today
Search results for all three names mix stale previews with invented spec sheets. Here is what first-party sources showed on August 10, 2026.
| SEEDANCE 2.5 | Released. Live on ByteDance Seed's official site; the API opened August 7, and platforms like Artlist and Segmind bill for it today. |
| MINIMAX H3 | Released. Launched July 31 per MiniMax's own blog, weights published days later under its Community License. |
| WAN 3.0 | Public beta only. Alibaba's official Wan account announced it August 6 for Model Studio and Qwen Cloud; access remains preview and invitation-only, with no general release and no weights. |
Ignore pages claiming Wan 3.0 shipped 4K output, 60B parameters, or April open weights; none of it survives contact with Alibaba's own channels. Pages calling Seedance 2.5 upcoming predate its launch week.
The full spec comparison
VERIFIED SPECS AND RATES AS OF AUGUST 10, 2026
| Spec | Seedance 2.5 | MiniMax H3 | Wan 3.0 |
|---|---|---|---|
| Developer | ByteDance Seed | MiniMax (Hailuo) | Alibaba Tongyi Lab |
| Released | Jul 31, 2026 (API Aug 7) | Jul 31, 2026 (weights days later) | Aug 6, 2026 (public beta only) |
| Max native clip | 30s one pass, extend twice | 15s, whole seconds only | 30s single continuous shot |
| Output today | 480p, 720p on public APIs | 768p and native 2K, 24 fps | 480p, 720p, 1080p tiers |
| Audio | Native, synced, audio can drive pacing and lip sync | Native stereo: dialogue, effects, score, voice transfer | Included, flagged by Alibaba as a weak point |
| Reference inputs | 50 files: 30 images, 10 videos, 10 audio | 12 files: 9 images, 3 videos, 3 audio | Text, media, plus doc, xls, ppt, pdf, md, web pages |
| Editing | Region-level edits, green screen, 3D white-model blockouts | Instruction-based local edits, ranked first for editing | Generation and editing folded into one model |
| Open weights | No | Yes, MiniMax Community License | Pledged Apache 2.0, not shipped |
| API list price | ~$0.24/s at 720p | $0.13/s at 2K | $0.20/s at 1080p ($0.05 at 480p) |
| Where it runs | Jimeng, Dreamina, CapCut, Volcano API, Krea, Artlist, more | Hailuo app, MiniMax API, fal, Krea, Segmind, your own GPUs | Alibaba Cloud Model Studio and Qwen Cloud beta only |
What each one is actually like to use
Seedance 2.5 ByteDance Seed · Jul 31, 2026
Max clip 30s native | Output 480p / 720p | References 50 files | Rate ~$0.24/s @720p |
WHERE IT WINS + Holds character, lighting, and motion across a full 30-second take with no stitching. + Takes 30 images, 10 video clips, and 10 audio tracks as references in one job. + A single music or voice track can drive pacing, beat matching, and lip sync on its own. + Region-level edits, green-screen swaps, and 3D white-model blockouts for staging shots before the model lights them. | WHERE IT FALLS SHORT − Public endpoints top out at 720p right now, whatever the 4K launch talk implied. − Runs roughly double the credit cost of Seedance 2.0 for the same settings on Dreamina. − No free API tier; the account needs a balance above $30 before the model switches on. − Long takes magnify mistakes. A bad brief burns 30 seconds of billing instead of five. | ||
MiniMax H3 MiniMax / Hailuo · Jul 31, 2026
Max clip 15s | Output 768p / 2K | References 12 files | Rate $0.13/s @2K |
WHERE IT WINS + Native 2K at 24 fps with synchronized stereo dialogue, effects, and score in one pass. + Edits you request in a sentence: swap a product, rewrite signage, relight day to night. + Renders clean on-screen text, subtitles, brand marks, and interface animations. + Open weights under the MiniMax Community License, plus a Turbo mode around 3.5x faster. | WHERE IT FALLS SHORT − The 15-second ceiling forces stitching for anything with a real narrative arc. − Hosted renders averaged a bit over four minutes in my queue, which slows iteration. − Local inference outputs 768p; reaching 2K needs a separate regeneration pass. − Reference video gets billed by its own duration on top of the output seconds. | ||
Wan 3.0 Alibaba Tongyi Lab · Aug 6, 2026
Max clip 30s one shot | Output Up to 1080p | References Media + docs | Rate $0.20/s @1080p |
WHERE IT WINS + Reads doc, xls, ppt, pdf, and md files, plus live web pages, as creative references. + One unified model covers generation, reference work, editing, and motion driving. + Director-grade camera language: pushes, pulls, pans, and tracking shots hold steady. + Cheapest full-length take of the three: a 30-second 1080p clip lists at $6. | WHERE IT FALLS SHORT − Public beta only, labeled preview and invitation-only, with no general release. − Alibaba itself flags audio quality and on-screen text accuracy as weak points. − The Apache 2.0 open-weights pledge has not shipped, and the Wan record on such pledges is patchy. − No independent benchmark scores yet, so quality claims rest on Alibaba's own demos. | ||
Who can hold a scene together
My coffee spot was the clearest test. Seedance 2.5 delivered the full arc in one take: beans, pour, steam, logo, done, with the barista's face consistent from first frame to last. Wan 3.0 also held a continuous 30-second shot, and its camera moves were the smoothest of the three, though only through borrowed beta access. H3 needed three clips and twenty minutes in an editor to tell the same story, and the cut between clips two and three never stopped looking like a cut.
EDGE SEEDANCE 2.5
Round 2 · Resolution reality check
The 4K claim vs what actually renders
Here is where marketing and shipping product split. Seedance 2.5 launch chatter referenced 4K, yet every public platform I generated on serves 480p and 720p, and Artlist openly points customers back to Seedance 2.0 when 4K matters. H3 delivers native 2K everywhere, and the extra pixels show in product close-ups: fabric texture, packaging edges, condensation on glass. Wan 3.0 tops out at 1080p across its three pricing tiers.
EDGE MINIMAX H3
Round 3 · Audio
Sound out of the box
Both frontrunners generate audio in the same pass as the picture, and neither charges extra. H3 outputs stereo with dialogue, foley, and score timed to the cut, and it can transfer a voice from a reference recording onto a character, which cut a dubbing step from my dialogue brief. Seedance 2.5 flips the relationship: hand it a music track or a voiceover and the track drives pacing, beat matching, and lip sync, with native support across 10+ languages. Wan 3.0 ships audio too, but Alibaba has been unusually honest that quality lags, and my beta clips confirmed it: serviceable ambience, mushy dialogue.
EDGE MINIMAX H3 · SEEDANCE 2.5 FOR MULTILINGUAL LIP SYNC
Round 4 · Reference control
How much of your world each model can hold
Reference capacity is the quiet spec that decides whether a model fits brand work. Seedance 2.5 accepts 50 files per job, enough for an entire brand kit: character sheets, product angles, motion clips, and sound cues all at once. H3 caps at 12 files, enough for every real brief I ran, and the first five images carry no extra charge. Wan 3.0 plays a different game entirely: it reads spreadsheets, slide decks, PDFs, and web pages as source material. Watching it turn a pricing sheet into a moving product explainer was the strangest moment of my testing week.
EDGE SEEDANCE 2.5 FOR VOLUME · WAN 3.0 FOR INPUT TYPES
Round 5 · Editing after the render
Fixing a shot without rolling the dice again
This round mattered more than expected, because the expensive part of AI video is regenerating a clip you almost liked. H3 currently ranks first on the Artificial Analysis video editing leaderboard, and my experience matched: I asked it to change a coffee cup's label in plain English and the edit landed on the label alone, with the rest of the shot untouched. Seedance 2.5 counters with region-level edits, green-screen background swaps, and 3D white-model blockouts that let you stage geometry before the model lights it. Wan 3.0 folds editing into its unified model, but with no third-party access, nobody outside the beta has stress-tested it.
EDGE MINIMAX H3
What a clip actually costs
Per-second rates hide the real story, so here is what my standard scenarios cost at list price. H3's numbers come from MiniMax's own rate card, Seedance figures from third-party API billing checked against my invoices, and Wan from the USD rates Alibaba published with the beta.
| Scenario | Seedance 2.5 (720p) | MiniMax H3 (2K) | Wan 3.0 (1080p) |
|---|---|---|---|
| 10s vertical clip | ~$2.40 | $1.30 | $2.00 |
| 15s ad spot | ~$3.60 | $1.95 | $3.00 |
| Full 30s take | ~$7.20 | Not possible in one pass | $6.00 |
| Cheap draft mode | 480p at roughly 45% of the 720p rate | 768p tier at $0.09/s, currently gated | 480p at $0.05/s |
Two costs the table cannot show: access and retakes. H3 runs through its own app, a half dozen third-party platforms, and your own GPUs under the license terms. Seedance 2.5 demands a funded account with no free API quota; on Dreamina an 8-second 720p clip drained 296 credits, about $3.57 on the entry plan and twice what Seedance 2.0 charges. Wan 3.0 posts the cheapest full-length take in the group and stays the hardest to reach, since the beta sits on Alibaba's own surfaces and no aggregator hosts it yet.
EDGE MINIMAX H3
Match the model to the job
| You are making | Use | Because |
|---|---|---|
| A 30s brand film with a story arc | SEEDANCE 2.5 | One continuous take, 50 references, consistent characters |
| Social clips at volume on a budget | MINIMAX H3 | $1.95 per finished 15s clip, sound included, retakes are cheap |
| Product video with readable labels or UI | MINIMAX H3 | Best on-screen text rendering of the three, at 2K |
| Multilingual talking-head or dubbed dialogue | SEEDANCE 2.5 | Audio-driven lip sync across 10+ languages in one pass |
| Video from a slide deck or PDF | WAN 3.0 | Only model that reads documents and web pages as input |
| Self-hosted generation on your own GPUs | MINIMAX H3 | Only one with weights you can download today |
| Long single-take camera choreography | WAN 3.0 | Director-grade camera moves across a full 30s shot |
| Fixing one detail in a clip you like | MINIMAX H3 | Sentence-level edits that land only where you point them |
Final verdict
When the comparison spreadsheet closed and real client work resumed, H3 was the tab still open. That surprised me, since on paper it loses the biggest number: 15 seconds against 30. In practice, most of what I ship is under 15 seconds anyway, and H3's combination of 2K output, sound that arrives finished, sentence-level edits, and $1.95 per maxed-out clip means I can afford to generate five takes and keep the best one. The four-minute queue is the tax, and I pay it.
Seedance 2.5 earned a different place in my workflow: the planned shot. When a brief needs one continuous scene, a stack of brand references, or dialogue synced in Mandarin and English from the same generation, nothing else here competes. I draft at 480p first, because a botched 30-second render at 720p costs more than lunch, and my first two coffee-spot attempts were exactly that kind of expensive lesson.
Wan 3.0 is the one I am watching rather than paying for. The document-to-video mode is a real idea, not a gimmick, and the 30-second single shot with confident camera work suggests the underlying model is strong. But an invitation-gated beta, audio its own maker apologizes for, and an open-weights promise the Wan line has broken before add up to a wait-and-verify position. If those Apache 2.0 weights actually ship, this entire ranking gets rewritten, and I will be first in the download queue.
Until then: H3 for the daily grind, Seedance 2.5 for the hero shot, Wan 3.0 for the watchlist.