Wan 3.0 API : Turn Text, Frames and References Into AI Videos up to 30 Seconds

The Wan 3.0 API lets your app generate videos with sound using the Wan 3.0 model over HTTP. On NanoAPI you send POST /v1/tasks with model wan-3.0, a version (standard or prime) and one of three modes (text_to_video, image_to_video or reference_to_video), then get the finished video by webhook or by polling.

You pay per second of the video you ask for: duration × the price per second of the version and resolution, and prime costs more than standard. All three modes cost the same, and reference images, videos, audio and generated sound don't change the price. The live prices are in the table above, on the pricing page and in GET /v1/products, and the studio shows the exact cost before you generate. Failed runs are not charged.

Both versions take exactly the same modes and parameters; prime is the higher tier and costs more per second. Try the same prompt with both in the studio above to see which one fits your use, then set variant in your API requests.

Yes. Sign in and use the studio at the top of this page: it runs the same Wan 3.0 model at the same prices as the API and draws on the same balance, so you can check results and costs before writing code. Top-ups start at $10.

From 2 to 30 seconds, set in whole seconds with duration, at 480p, 720p or 1080p. Text to video takes 16:9, 9:16, 1:1, 4:3 or 3:4; reference to video takes the same or auto, which follows your reference images and videos (16:9 with audio only); image to video takes no aspect_ratio and follows the first frame. variant, mode, prompt, resolution and duration are required on every request.

In reference to video, send up to 10 image_urls, up to 5 video_urls (each 1–15 seconds, 15 seconds in total, and together with duration at most 30 seconds) and up to 5 MP3 or WAV audio_urls (15 seconds in total). Send at least one kind; audio alone works. In the prompt, refer to each file by type and order as <<<image_1>>>, <<<video_1>>> or <<<audio_1>>> to say which character appears, which camera move to follow or which track to play.

Videos come with generated sound unless you set generate_audio to false, and the price is the same either way. The video with sound is at outputs[].url, and preview_url is a compressed copy without sound. enable_thinking is optional: set it to true and the model reasons about your prompt before generating, so the video follows it more closely, but the task takes longer.

Most videos take 1 to 5 minutes; higher resolutions, longer videos and enable_thinking take longer. In production, pass a webhook_url to be notified once when the task finishes instead of polling. If a generation fails, the amount reserved for it goes back to your balance in full, and the task's error.type tells you why, for example a file that can't be downloaded or a reference video longer than 15 seconds. Finished videos are kept for 15 days.