Wan vs Seedance

Top of the Board, by a Small Margin

Wan 3.0 leads Artificial Analysis's text-to-video-with-audio leaderboard with an Elo score of 1173, followed by Utopai X at 1166, Seedance 2.5 and Black Forest Labs' FLUX 3 Max at 1159, and MiniMax H3 at 1158. The top five are within 15 points of each other, a narrow spread for a blind-vote ranking. The leaderboard tells you that both models are at the frontier as of October 2026, not that one is clearly better for your shots. Kling 4.0 and other October releases hadn't been rated yet. Test both on your own prompts, and see Seedance vs Kling for the other big Chinese contender.

Inputs and Control

The two models take direction differently. Alibaba pitches Wan 3.0 as generating video "from any input": besides text, images, audio and video, it accepts documents such as PDFs, slide decks, spreadsheets and Markdown files. That makes it natural for turning existing material, such as a product sheet or a presentation, into a video without first writing a prompt by hand.

Seedance 2.5 is built for consistency across many assets. It accepts up to 50 references in one generation (30 images, 10 videos and 10 audio clips), four times Seedance 2.0's limit of 12, and offers director-style multi-shot camera control. For a commercial with a recurring character, a specific product and a set soundtrack, that depth of reference is a real advantage.

Audio

Both generate sound with the picture, but Seedance is further along on its own account. Seedance 2.5 generates native audio with lip-synced speech. Wan 3.0 produces audio with multilingual voice, but Alibaba itself has said the audio texture is "not yet where we want" it. If dialogue and lip-sync carry the scene, Seedance is the safer choice today; if you plan to replace or mix the audio anyway, it matters less. See text-to-speech for separate voice generation.

Access, Price and Regions

Wan 3.0 is the easier model to build a product on. Alibaba Cloud Model Studio publishes per-second list prices for international customers: $0.05 at 480p, $0.10 at 720p and $0.20 at 1080p for the standard model, with a Prime variant at $0.068 to $0.28 per second, and was offering a 30% discount on the standard model at the time of writing. Model Studio has endpoints in Singapore, Frankfurt, Virginia, Beijing and Hong Kong, though not every variant is offered in every region, which gives some choice over where requests are processed (see data sovereignty). Runway also hosts Wan 3.0.

Seedance 2.5 reaches far more consumers. It launched on Dreamina, along with Jimeng and Doubao in China, and reached CapCut's desktop app for Pro and Ultra members on August 7, 2026, billed per second in credits. Developers can use BytePlus ModelArk or third-party hosts, but resolutions and prices vary by host: the main API hosts offer 480p and 720p, and published per-second prices differ several-fold. Budget against the host you will actually use.

Openness: Different Lineages

Neither flagship has open weights. That's worth stating plainly for Wan, because some sites have claimed Wan 3.0 weights were released; Alibaba hasn't published them, and its Hugging Face organization has no repository for Wan 2.5, 2.6, 2.7 or 3.0. The difference is history. Alibaba released Wan 2.1 (February 2025) and Wan 2.2 (July 2025) as open weights under Apache-2.0, and has since published specialized open models such as Wan2.2-Animate-2 for character animation. Teams that want a self-hosted fallback can run Wan 2.2 today, at lower quality and shorter length; see LTX vs Wan. No open-weight Seedance model was found. See open-weight models.

Rights and Risk

Seedance carries a documented reputational risk. Within days of Seedance 2.0's February 2026 release, Disney, Paramount Skydance and Netflix sent ByteDance cease-and-desist letters; the Motion Picture Association and SAG-AFTRA condemned the model; and US senators wrote to ByteDance's CEO. ByteDance said it respects intellectual property and strengthened safeguards, and its global CapCut rollout of 2.0 blocked real faces, added watermarks and excluded the US. No comparable dispute involving Wan was found in the sources reviewed. Whatever model you use, avoid prompting for protected characters and real people without rights and consent, and disclose synthetic content where platforms or law require it (see AI content labeling).