视频生成模型的竞赛正在换赛道：上半场拼谁生得长，下半场拼谁生得便宜。8月24日通义万相3.0全面上线，两天后的8月26日，其Prime版本正式上架阿里云百炼——30秒长视频，开始按秒计价。
30秒，和“万物皆素材”
规格层面，万相3.0把单次生成时长拉到30秒，是上一代的两倍，输出帧率30fps，并原生输出台词、BGM与音效。更值得注意的在输入端：除了文、图、音视频，它支持直接传入doc、xls、ppt、pdf等文档和公开网页链接，让模型读完全文再生成视频，单次最多可参考20个多模态素材。把“读材料”和“出片子”合并成一步，这才是“文档直出视频”的完整含义。
价格表上见真章
国内API原价为480P 0.3元/秒、720P 0.6元/秒、1080P 1.2元/秒；国际站定价0.05、0.10、0.20美元/秒——一段30秒的1080P视频约6美元，480P草稿版只要1.5美元。按国际站定价口径，这大约是竞品Seedance 2.5的一半。当分辨率、时长、音效都被拉齐之后，价格成了最直观的那根杠杆。
双线赛马的阿里
在百炼的视频模型矩阵里，happyhorse-1.1与万相3.0系列并列出现在官方推荐位——同一个战场，两条产品线。叠加阿里8月24日完成约百亿美元级的配售，AI基础设施的弹药已经备足。对阿里来说，万相走旗舰对标，HappyHorse走差异化路线，赛马机制在AI视频上重演。
需要提醒的是，各家定价口径（分辨率、计费单位、是否含音效）并不完全对齐，跨平台比价前先看清脚注。但趋势已经清楚：时长战后是价格战，受益的终究是拿着创意排队的人。明天见。
The AI video race is switching tracks: the first half was about who could generate longer clips, the second half is about who can generate them cheaper. Wan 3.0 went fully live on August 24, and two days later, on August 26, its Prime version landed on Alibaba Cloud's Bailian platform — 30-second videos, now billed by the second.
30 seconds, and "everything is material"
On specs, Wan 3.0 stretches a single generation to 30 seconds — twice the previous generation — at 30fps, with native dialogue, BGM and sound effects. The input side matters even more: beyond text, images, audio and video, it accepts documents (doc, xls, ppt, pdf and more) and public web links, reading the material in full before generating the video, with up to 20 multimodal references per run. Merging "reading the material" and "producing the clip" into one step is what document-to-video actually means.
The price list does the talking
Domestic list API pricing is RMB 0.3/0.6/1.2 per second for 480P/720P/1080P; international pricing is $0.05, $0.10 and $0.20 — a 30-second 1080P clip runs about $6, and a 480P draft just $1.50. By international list pricing, that is roughly half of rival Seedance 2.5. Once resolution, duration and audio are all matched, price becomes the most visible lever.
Alibaba's two-track lineup
In Bailian's video model matrix, happyhorse-1.1 sits alongside the Wan 3.0 family in the official recommendations — two product lines on the same battlefield. Add Alibaba's roughly $10 billion share placement completed on August 24, and the ammo for AI infrastructure looks fully stocked. Wan takes the flagship head-on while HappyHorse differentiates; the horse-racing mechanism replays in AI video.
One caveat: pricing terms (resolution, billing units, audio inclusion) are not fully aligned across vendors, so read the footnotes before comparing. But the trend is clear — after the length war comes the price war, and the winners are the creators queuing with ideas. See you tomorrow.
当30秒视频按秒计价、价格只有对手一半，AI视频的竞争就从实验室竞赛变成了成本竞赛。
—— Dawn Vision编辑部
When a 30-second video is billed by the second at half a rival's price, the AI video race turns from a lab contest into a cost contest.
— The Dawn Vision Editorial Desk
本文以8月26日Prime上架百炼为时效锚点，8月24日的GA作为背景信息交代；价格均按API原价口径呈现（国内0.3-1.2元/秒、国际站0.05-0.20美元/秒），“约为Seedance 2.5一半”依据国际站定价对比并注明口径；全文未使用任何“首创”类表述；HappyHorse与万相并列的信息来自百炼官方模型矩阵页。
This piece anchors timeliness on the Prime launch on Bailian on August 26, with the August 24 general availability as background. All prices are presented at list API rates (RMB 0.3-1.2/second in China; $0.05-$0.20/second internationally), and the "roughly half of Seedance 2.5" comparison follows international list pricing with the basis noted. No "first-ever" claims are made; the HappyHorse-Wan parallel comes from Bailian's official model matrix page.
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理当日采集信号并交叉验证 4 个核心来源后生成，经编辑部人工审核。素材来源：Alibaba Cloud官方博客（规格与国际站定价）、阿里云帮助中心万相3.0 API参考（时长/帧率/输入能力）、阿里云百炼模型矩阵页（HappyHorse与万相并列）、千问AI平台模型定价页（国内API定价）。
This article was generated by the Dawn Vision cognitive engine from signals collected that day and cross-verified against 4 core sources, with human editorial review. Sources: Alibaba Cloud's official blog (specs and international pricing), Alibaba Cloud Help Center's Wan 3.0 API reference (duration, frame rate, input capabilities), Bailian's official model matrix page (HappyHorse-Wan parallel), and the Qianwen AI platform's model pricing page (domestic API pricing).
