The 30-Second Threshold: Inside Seedance 2.5, MiniMax H3, and Wan 3.0

Three AI video models arrived within twenty five days of each other and quietly rewrote what a single prompt can produce. Here is what each one actually does, what the benchmarks say, and what every second of output costs.

For most of the short history of AI video, one limit defined the medium: clips lasted a handful of seconds. A product ad needed stitching. A story needed cuts. A complete scene, with a beginning, a middle, and an end, simply did not fit inside one generation.

Between July 31 and August 24, 2026, that limit fell. ByteDance shipped Seedance 2.5 with 30-second single-pass generation and synchronized audio. MiniMax announced H3 the very same day, then published its weights, making a frontier-grade multimodal video model downloadable. Alibaba followed with Wan 3.0, matching the 30-second ceiling, adding document-to-video generation, and debuting at the top of the most watched community leaderboard.

This guide covers all three models in depth: the confirmed specifications, the honest caveats, the current benchmark standings, and verified pricing as of August 26, 2026. Every claim below traces back to launch documentation, official rate cards, or independent reporting.

How This Guide Was Built: The DRIVE Framework

Rather than lining up marketing claims, each model was assessed through five practical lenses, together called the DRIVE framework:

• Duration and continuity: how long a single generation runs, and how well it holds a scene together.
• References and control: what inputs the model accepts and how precisely creators can direct it.
• Integration and access: where the model actually runs, from consumer apps to APIs to local GPUs.
• Value: published first-party pricing, converted to cost per second of usable output.
• Ecosystem and licensing: open weights, regional restrictions, and commercial terms.

Sources include official announcements from ByteDance, MiniMax, and Alibaba Cloud, published rate cards from BytePlus ModelArk and the MiniMax platform, independent reporting from TechNode and The Decoder, and blind-vote rankings from the Artificial Analysis Video Arena. Where a number is provider-dependent or still unconfirmed, the text says so plainly.

Seedance 2.5: The Reference Powerhouse

Seedance 2.5 is ByteDance's flagship video model, previewed at the Volcano Engine conference in late June and released on July 31, 2026, with a public developer API following on August 7. It reached creators first through Jimeng AI and the Pro tier of Doubao, with API access through Volcano Engine Ark in China and BytePlus ModelArk internationally. It is also available online through platforms such as Seedance 2.5 generator.

The headline capability is a coherent 4 to 30 second clip generated in one pass, with dialogue, music, and sound effects produced together with the visuals rather than bolted on afterward. Multi-turn extension lets strong clips continue beyond that window while preserving lighting, action, and context.

Where Seedance 2.5 stands apart

• Up to 50 reference assets in a single job: as many as 30 images, 10 video clips, and 10 audio files, used to lock characters, products, locations, motion, and sound.
• Second-level timestamp direction, so a 30-second idea can be divided into precise intervals with their own shots, actions, and audio cues.
• Storyboard, keyframe, and previs inputs that define shot order and staging before generation begins.
• Local segment editing, which revises a selected section while keeping the rest of the video stable.
• Multilingual dialogue and subtitles across more than 10 languages, with improved lip synchronization.

The honest caveats

Some creator products market 4K within the Seedance family, but independently verified API routes currently bill at 480p, 720p, and 1080p. Teams budgeting an API pipeline should treat 4K as a creator-product tier rather than a guaranteed endpoint. Seedance 2.5 also remains fully closed: there are no weights to download and no announced plans for a local version. Hands-on reviews describe the jump from Seedance 2.0 as a refinement in motion handling and prompt adherence rather than a dramatic quality leap, with the longer duration and expanded references doing most of the heavy lifting.

MiniMax H3: The Open-Weight Contender

MiniMax H3 is the model behind the Hailuo line, announced on July 31, 2026 and published as open weights on Hugging Face on August 3. It is a 33-billion-parameter omni-modal system: text, images, video, and audio all enter one shared context, and the model produces video with native stereo audio, up to 15 seconds long, at up to 2K resolution and 24 frames per second. It can be tried online through hosts such as Topview's MiniMax H3 page as well as fal.ai and the Hailuo app.

What builders get

• Text-to-video, image-to-video, first-and-last-frame control, and reference-to-video, with hosted endpoints accepting up to 9 images, 3 video clips, and 3 audio files per job.
• Instruction-led video editing and V2V motion transfer, which moves timing, camera language, and performance from a guide video into a new subject or style.
• Strong text and brand rendering, a rare strength for logos, packaging, and on-screen typography in commercial work.
• Day-one ComfyUI support, plus INT8 and NVFP4 quantizations and deployment paths through vLLM, SGLang, and diffusers.

Read the license before self-hosting

The open release covers the H3-Base checkpoints (one for text and frame-guided generation, one for reference-based generation), while the Context-IR preprocessing layer that powers the full hosted pipeline remains closed. More importantly, the MiniMax H3 Community License excludes local deployment in the United States, the European Union, the United Kingdom, and South Korea unless MiniMax grants separate authorization, and it forbids training smaller models on H3's outputs. In practice, H3 is best described as regionally available open weights rather than a fully unrestricted release. Teams in excluded regions can still use hosted APIs.

On quality, the community verdict has been emphatic: H3 leads every open-weight bracket of the Artificial Analysis arena, ranks near the very top of the overall text-to-video board, and holds the top position in audio-inclusive video editing. For an openly downloadable model, that is unprecedented territory.

Wan 3.0: The Document-to-Video Pioneer

Wan 3.0 comes from Alibaba's Tongyi Lab. A public beta opened on August 6, 2026 through Alibaba Cloud Model Studio and Qwen Cloud, and the general release landed on August 24, one day after Alibaba closed a share placement of roughly 10.2 billion US dollars earmarked entirely for AI. Access runs through Alibaba's hosted services and partner platforms.

Like Seedance 2.5, Wan 3.0 generates up to 30 seconds in a single pass, double the 15-second ceiling of its predecessor Wan 2.7, at resolutions up to 1080p with audio generated by default. Alibaba reports improved instruction following, better consistency across shots, and higher audio quality, and cites real deployments across short drama, advertising, tourism promotion, and music videos.

The feature no rival matches

Wan 3.0's signature capability is source-material input. A single request can include not just text, images, video, and audio, but also documents in DOC, XLS, PPT, PDF, TXT, Keynote, Pages, Numbers, and Markdown formats, plus public webpage URLs. The model reads the content and builds a video sequence from it, turning a pitch deck, a report, or a product page directly into motion. Among current mainstream video models, this document-to-video and webpage-to-video path is unique.

• Smart duration recommends a clip length that fits the density of the brief, so short ideas are not padded to fill time.
• Omni-Reference holds a character, product, or set consistent across every shot of a sequence.
• An extension capability stretches a narrative timeline beyond the initial generation.

The trade-off

Wan 3.0 is closed. Alibaba built its video reputation on open releases, but the last open-weight Wan flagship remains Wan 2.2 from July 2025, and no Wan 3.0 files exist on Hugging Face, GitHub, or ModelScope. Any site advertising a Wan 3.0 download is not offering the real model. The hosted API also remains in preview, with approval required and endpoints documented in Beijing, Singapore, and Virginia.

Side by Side: The Full Specification Picture

The table below condenses the verified specifications of all three models as of August 26, 2026.

Table: Confirmed specifications from launch documentation and provider rate cards, checked August 26, 2026.

Duration: The New 30-Second Standard

Clip length is the clearest single number separating this generation from the last. A 30-second window covers a complete broadcast spot, a full product story, or a scene with an actual arc, all without stitching. Seedance 2.5 and Wan 3.0 both reach it in one pass. MiniMax H3 stops at 15 seconds, though its 2K output resolution is the highest of the three.

Figure 1: Maximum single-pass clip length. Gray bars show the previous generation and a leading rival for context.

What the Benchmarks Actually Say

The most credible public quality signal in AI video is the Artificial Analysis Video Arena, where users vote between two unlabeled clips generated from the same prompt. As of late August 2026, Wan 3.0 leads the text-to-video board for models with audio at an Elo of 1241, narrowly ahead of Google's Gemini Omni Flash at 1237 and MiniMax H3 at 1226.

Figure 2: Artificial Analysis text-to-video arena, models with audio, late August 2026.

Three caveats keep this honest. First, Seedance 2.5 does not yet appear on the public board, so its predecessor Seedance 2.0 stands in for the family; its own ranking is still forming. Second, gaps of a few Elo points sit within normal voting noise, so the top three are better read as a cluster than a podium. Third, image-to-video tells a different story: there, the Seedance family holds the top audio-inclusive spot, with H3 close behind. Rankings on these boards move week to week, so a live check is always worth thirty seconds before a big production decision.

Verified Pricing: What Each Second Costs

Published first-party rates, checked on August 26, 2026, are the fairest baseline, since third-party hosts add their own margins and the same generation can cost anywhere from ten cents to over a dollar per second depending on the route.

Figure 3: First-party list rates per second of generated video. Seedance 2.5 figures are derived from BytePlus ModelArk's published 5-second examples.

• Wan 3.0 lists $0.05, $0.10, and $0.20 per output second at 480p, 720p, and 1080p, with a temporary 30 percent API discount running from August 24 to September 23, 2026 on selected platforms.
• Seedance 2.5 bills by tokens on BytePlus ModelArk ($10.70 per million output tokens without video input). The published 5-second 16:9 examples work out to roughly $0.10 per second at 480p and $0.23 per second at 720p. Third-party per-second routes span roughly $0.10 to $0.43 depending on provider and tier.
• MiniMax H3 costs $0.13 per second at 2K on MiniMax's own platform, with a cheaper 768p tier around $0.08 per second in limited availability. Generated audio adds no extra charge on any of the three.

In practical terms: a full 30-second 1080p Wan 3.0 clip lists at $6.00, or $4.20 during the launch promotion. A 30-second 720p Seedance 2.5 clip lands near $6.94. H3's 15-second maximum at 2K costs $1.95. The real budget driver, though, is rerolls: a model that needs five attempts per accepted clip is more expensive than its rate card suggests, so cost per accepted output is the number worth tracking.

Matching the Model to the Job

• Reference-heavy brand and character work: Seedance 2.5. Fifty locked assets, timestamp direction, and segment editing make it the strongest choice when identity and product fidelity across a 30-second story are non-negotiable.
• Builders, researchers, and private pipelines: MiniMax H3. Downloadable weights, ComfyUI support, motion transfer, and top-tier editing, provided the deployment region is covered by the license or a hosted API is acceptable.
• Turning existing material into video: Wan 3.0. Decks, PDFs, spreadsheets, and product pages become explainers and campaign films without a manual rewrite of the brief.
• The tightest budgets: Wan 3.0 again, whose $0.05 per second 480p tier makes it the cheapest way to draft and iterate before committing to a high-resolution render.
• The sharpest pixels: MiniMax H3, the only one of the three with native 2K output today.

Trying All Three in One Place

Each model lives on a different first-party platform, which makes side-by-side testing tedious. Multi-model platforms solve this: Topview, for example, runs workflows for Seedance 2.5, MiniMax H3, and Wan 3.0 under one roof, alongside storyboard tools, previs, and conversational editing, so the same prompt and references can be tested across models before a campaign locks in. Whichever route a team picks, the sensible first step is the same: run one identical brief through all three and judge the results against the actual deliverable, not the demo reel.