Editorial Team
5 min read
Wan 3.0 Arrives on fal, but Alibaba’s Direct API Has the Lower Promotional Price
Alibaba Tongyi Lab’s Wan 3.0 video model is available through fal with up to 30-second, 1080p audio-video generation and reference inputs. The launch broadens access, but Alibaba Cloud Model Studio’s stated 30% promotional discount makes direct access cheaper at the published rates through September 24, 2026.
- Wan 3.0 is available on fal for text-, image- and reference-to-video generation of two to 30 seconds, at up to 1080p with audio generated in the same pass.
- fal lists $0.05, $0.10 and $0.20 per generated second at 480p, 720p and 1080p; Alibaba Cloud lists a 30% promotional reduction on those rates through September 24, 2026.
- The source material supplies product demonstrations and configuration claims, not a common performance test against rival video models or between the two access routes.
Wan 3.0, the next generation of Alibaba Tongyi Lab’s Wan video-model family after Wan 2.1 through 2.7, is available through fal’s hosted API. fal’s product page says the commercial model can make a single two-to-30-second video generation from text, a still image or supplied reference material, with audio generated alongside the picture.
The addition gives developers another route to an Alibaba model that is also generally available directly in Alibaba Cloud Model Studio. It does not, on the retained evidence, establish that fal is cheaper or that Wan 3.0 outperforms its named alternatives. In fact, Alibaba’s current published promotion gives the direct route a lower stated per-second price until its advertised expiry.
fal’s published Wan 3.0 text-to-video product example. Source: fal Wan 3 overview.
Alibaba Cloud Model Studio’s company-reported Wan 3.0 reference-to-video showcase pairs three inputs with a 24-second output example. Source: Alibaba Cloud Model Studio Wan 3.0.
The substantive new capability is a longer, conditioned generation
fal says Wan 3.0 supports 480p, 720p and 1080p output, five aspect ratios, and audio by default. Its optional “thinking” mode is described as reasoning about composition, staging and motion before rendering. Those are product claims, not independently measured results.
The more specific workflow addition is the reference interface. A request can contain up to 10 images, five video clips totalling 15 seconds at 16 frames per second or more, and five audio tracks totalling 15 seconds. With thinking enabled, a document or public web page can also be an input. That gives an editor a way to supply character, location, product or audio references rather than relying on a text prompt alone; it does not prove that the model will preserve them reliably.
