← Back to blog
AiAbout 5 min read

Baidu Yijing pushes video generation to 0.2 yuan a second, and the digital human business gets rewritten first

Published Nov 14, 2026
Baidu Yijing pushes video generation to 0.2 yuan a second, and the digital human business gets rewritten first

On November 12 the Baidu World 2026 opening video showed two people, one Robin Li and another Robin Li. The host, Xijiajia, a digital human herself, questioned the "real" one; the camera cut, and the "digital Robin Li" answered from the same sofa and lamp. The applause ran half a beat longer than for anything else.

Baidu Yijing generated the clip. Yijing is Baidu's digital human line, upgraded with the conference. The three official numbers are blunt: video generation down to 0.2 yuan per second, 8 to 9 micro-expressions per second under realistic control, full-frame real-time interaction under 2 seconds.

Of the three, the first stings most.

What 0.2 yuan a second actually buys

That month, ByteDance's Seedance 2.5 Flash launched on November 12, with 720p text-to-video and image-to-video at about 0.99 yuan per second, 34% cheaper than the flagship and twice as fast. Google's Gemini Omni 1.2 Flash opened on November 5, quoting 720p at 0.07 dollars a second, 30% below the last generation.

0.2 yuan is roughly 0.028 dollars, an order of magnitude below general video models. The job differs. Yijing makes talking-head material: fixed camera, fixed person, one face called up again and again. It won't give cinematic shots on request. The price gets this low because the output range was deliberately narrowed.

The price of cheap is written in the word "acting"

Digital humans have been stuck on one point for years. The face convinces, the performance does not. Audiences forgive a slightly off face; they won't forgive a digital human whose eyebrows keep smiling through a sad passage.

Yijing splits micro-expressions into 8 to 9 control points per second. It also released a Wenxin Yijing micro-expression model, live that day on Qianfan with an API for enterprise customers. More expressions alone would mean little. What matters is tunable parameters: the pipeline runs a first pass, so nobody picks takes by hand.

Two identical faders side by side on a mixing console, one sitting slightly out of line with the other

A count of expressions is not a performance. An eyebrow switching nine times a second and an actor who knows when to leave it alone are separated by judgment. Resolution has nothing to do with it. No public evaluation data exists, only a line about emotional value. Real results wait for customers to publish footage.

The Short Comic Drama Studio is the actual product here

Yijing also opened a Short Comic Drama Studio, covering idea, script, storyboard and final cut. One line in the release is easy to skip: creators can take a "work already validated" and, in one click, localize plot, characters and settings for an overseas market.

The weight sits on "already validated". It admits something: the tool won't find you a good story. It takes a story that worked at home and re-skins it for another market. Swap faces, streets, language, with cost counted by the second. A capacity tool, some distance from a creation tool.

The three costliest items in overseas short drama these two years are actors, locations and localized reshoots. Virtual production pushed the first two down. The third has not moved. Studio aims straight at it.

Another layer upstream: Ciyuansuo

The same event launched a model marketplace called Ciyuansuo. Robin Li's framing: a new model appeared almost daily for nine months, token relay stations answer for neither quality nor stability, and more models mean harder choices. Ciyuansuo sells verification and credit: fingerprint comparison and capability tests, plus spot checks and full-chain monitoring.

One month in: over 100 supplier resources onboarded, daily calls past 50 million, daily tokens above 700 billion.

Set that beside 0.2 yuan a second and the logic holds. Video generation costs GPU hours. It also costs the trial and error of confirming whether an endpoint serves the model it claims. Every middleman adds cost. Ciyuansuo wants to standardize that uncertainty.

Robin Li was talking about organization, and models came second

The keynote ran on "intelligent organization". Three principles: break job boundaries, let tacit knowledge settle, design closed loops. The formula was "measurable + looping = continuous evolution". His analogy was the industrial era: swap the motors, keep the layout, and the expected gains never arrive.

Internal figures came out as evidence. Daily active agents rose from 5,000 to 37,000. Effective tasks per agent week rose from 84,000 to over 1 million. The company built 4,300 Workspaces and 20,000 organizational Skills.

Talk and launch fit together. 0.2 yuan a second is cheap only inside an organization shipping hundreds of clips a day. Make three films a month and the savings won't cover the people needed to wire it up.

Who should move now

Three groups are worth a look. Teams doing overseas short drama and localized dubbing, where reshoots are the easiest line to replace. Content operations in e-commerce and local services, where talking-head material is a consumable and per-second pricing favors you. Service providers building digital employee integrations, since Qianfan's API lets you cost a single clip.

Brand advertising and film-grade projects can wait. 0.2 yuan doesn't buy camera language, or an actor's hesitation. The two-Robin-Li video works half because Li agreed to play along, sitting there while another version of him spoke. That arrangement was a creative decision. The model had little to do with it.

Related articles

Google replaced the model behind Search, and the number that moved was a shell benchmark

Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.

Three ways a coding agent wastes money, measured on 30 repositories

TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.

earshot gives coding agents 922 sounds, and every one is rendered from code

An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.

The agents stealing API keys were not sophisticated. The gateways were open.

Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.