Baidu opens ERNIE Wizshot-ME video API at 0.7 yuan a second
Baidu's Yijing digital-human unit showed its own video model at as little as 0.2 yuan a second, and Baidu calls the Wizshot-ME price half the market rate.

Key takeaways
- Baidu opened API access to ERNIE-Wizshot-ME, a talking-character video model, on its Qianfan cloud platform on November 12, 2026, at 0.7 yuan per second of generated video.
- Baidu's own product page says a five-second talking video costs about 3.5 yuan, and at the same rate a 15-second clip comes to 10.5 yuan.
- Baidu's Yijing digital-human business showed its own video generation model at as little as 0.2 yuan a second, with a fast version that returns 10 seconds of video in 10 seconds.
- No independent price or quality measurement of either model existed as of November 13, 2026, and the claim that Wizshot-ME costs half the market rate is Baidu's own.
Baidu opened API access to ERNIE-Wizshot-ME, a model that generates short videos of people speaking, on its Qianfan cloud platform on November 12, 2026, at 0.7 yuan per second of generated video, alongside a second, cheaper video model from the company's digital-human unit priced at as little as 0.2 yuan a second.
Baidu introduced both at Baidu World 2026, its annual conference in Beijing, according to Baidu's release. ERNIE-Wizshot-ME, whose Chinese name is 文心一镜微表情大模型, is sold to enterprise customers and partners through the Qianfan API. Qianfan's billing page lists the model at 0.7 yuan for each second of delivered video, and Baidu's product page says a five-second talking video costs about 3.5 yuan. At the same rate a 15-second clip comes to 10.5 yuan. Converted at 6.685071 yuan to the dollar, the rate published on November 13, 2026, 0.7 yuan is about 10 cents a second and 0.2 yuan is about 3 cents.
Wizshot-ME processes text, images and speech together and takes scripts of up to 4,000 Chinese characters, Baidu's release says. Qianfan's product listing describes 1080p output and clips of four to 15 seconds, as AI Ping reports. The create-task documentation sets a stricter internal rule: the synthesized speech and any silence in a request must total four to 16 seconds as measured by the server, or the task fails.
Each request defines one to 10 personas, and Baidu suggests three or fewer. A persona needs a portrait image in JPEG, PNG or WebP, between 300 and 5,000 pixels per side and up to 10 MB, plus a WAV voice reference of 10 to 60 seconds that contains speech. A request carries up to 10 shots, with Baidu suggesting four or fewer; each shot needs a storyboard image whose short side is at least 720 pixels, a camera direction and the names of the characters on screen. Dialogue is supplied per shot with an emotion value, optional performance notes such as a nod on a keyword, and voice settings for pace and tone.
Generation is asynchronous. The create call returns a task ID, and a separate query request reports the job as pending, processing, succeeded or failed. A finished job returns a video link that is valid for 24 hours; Baidu keeps the video for up to 30 days. Qianfan's model list sets Wizshot-ME at one concurrent task and one queued task, which constrains teams submitting several scenes in a row.
Baidu's release says Wizshot-ME "leads the industry for podcasts and short animated dramas at half the typical cost," and the company's presentation put the price at 50% below the market. That is Baidu's own claim. The fast version generates about a minute of video in a minute, the company says.
At the same event, Baidu Yijing (百度一镜), the company's digital-human agent, showed its own video generation model at as little as 0.2 yuan a second, per Baidu's release and Xinhua. The fast version returns 10 seconds of video in 10 seconds, renders eight to nine facial micro-expressions a second and answers in under two seconds, Xinhua and neurocartel report. Yijing also introduced a short animated drama Studio that covers concept, script, storyboard and final cut, with one-click localization for other markets. Both video models are separate from the ERNIE 5.5 Flash Preview language model Baidu showed the same day.
The nearest published per-second price from another Chinese lab is BytePlus's, whose own pricing page lists Dreamina Seedance 2.5 Flash at $0.07 a second at 480p and $0.15 a second at 720p, with 1080p unsupported, as of November 13, 2026. Seedance 2.5 is a general text-to-video model rather than a talking-character model, so the two are not like for like, but on list price Wizshot-ME's roughly 10 cents a second falls between Flash's two rates, and Yijing's roughly 3 cents sits below both.
No independent price or quality measurement of either model existed as of November 13, 2026. Access runs through a Qianfan API key, the documentation and examples are in Chinese, and Baidu has not said how easily accounts outside mainland China can sign up. If the rates hold, the cost of a talking-character retake falls to a few cents, and the scarce input becomes the script and the shot list rather than the render budget.
Baidu has not said whether the 0.2 yuan Yijing rate will reach the Qianfan API, which lists ERNIE-Wizshot-ME at 0.7 yuan a second.
Sources
- cloud.baidu.com - Qianfan billing page: ERNIE-Wizshot-ME listed at 0.7 yuan per second of generated video
- cloud.baidu.com - create-task API doc: personas, portrait and voice-reference specs, shots, the 4-16 second dialogue rule
- cloud.baidu.com - query API doc: asynchronous task, 24-hour video URL, 30-day storage
- cloud.baidu.com - Baidu product page: a five-second talking video about 3.5 yuan
- us.acrofan.com - Baidu press release: Wizshot-ME specs, 4,000-character scripts, "half the typical cost", Yijing RMB 0.2 per second
- xinhuanet.com - Xinhua: Yijing 0.2 yuan per second, 8-9 micro-expressions a second, two-second response, short-drama Studio
- aipingnews.com - AI Ping: 1080p output, 4-15 second clips, inputs, one concurrent job
- llms.blog - LLMs Blog: announced by Jia Lei, no open weights, no statement on accounts outside mainland China
- neurocartel.com - neurocartel: Yijing fast version 10 seconds in 10 seconds, the 50% claim, no independent measurement
- docs.byteplus.com - BytePlus's own pricing page: Dreamina Seedance 2.5 Flash at $0.07-$0.15 per second
- open.er-api.com - exchange rate used for the dollar figures: 6.685071 yuan to the dollar, read November 13, 2026