Wan 3.0
Wan 3.0
One-line summary: Alibaba's August 2026 general video model — 30 s single pass, document-to-video, official API $0.05–$0.20/s — not yet on the Artificial Analysis tables fetched 2026-08-18.
What it is
Public-beta successor to wan-2-7 on Alibaba Cloud Model Studio / Qwen Cloud. Official 7 Aug 2026 announcement: up to 30 s per clip (vs Wan2.7's 15 s), multimodal references including web pages and office documents, and stricter character/product/space/style lock. The 13 Aug 2026 product post adds Model Studio list prices and a "everything to video" pitch. From 2026-08-18-autoresearch-best-ai-video-generation-tools.
Why it matters to ai-video-generation
Longest official single-generation hosted clip in this ingest, and the first Wan line that eats PPT/PDF as input. Reference consistency and "no more same-face AI people" are vendor claims aimed at the same identity problem as wan-animate, but this is not a pose-driven Animate drop and has no AA Elo yet.
Key facts
- Vendor: Alibaba Cloud Model Studio.
- Max duration: 30 s, plus "smart duration" and extension.
- Inputs: text, image, audio, video, and one document/link (doc/xls/ppt/pdf/txt/key/pages/numbers/md; ≤100 MB, ≤50 pages).
- Output: 480P / 720P / 1080P.
- Official API: $0.05/s (480P), $0.10/s (720P), $0.20/s (1080P). A 30 s 1080P generation is $6.00.
- Authors flag weak audio texture and on-screen text as current limits.
Strengths (from our perspective)
- Length and document input are unique among official pages fetched this run.
- 720P list price matches Omni Flash's ~$0.10/s; 480P drafts are cheaper.
Weaknesses (from our perspective)
- Beta; no independent Elo on the AA pages fetched 2026-08-18 — do not rank it above gemini-omni-flash / minimax-h3 / seedance-2-0 on quality.
- Closed hosted API.
- Audio/text rendering admitted-weak.