Of the 128,000 short dramas published in China in the first quarter of 2026, 95% were produced as AI video, on the count kept by the China Netcasting Services Association. The sector directly employed 690,000 people, and fifteen million Chinese workers name livestreaming as their main occupation.
Key Takeaways
- 128,000 short dramas shipped in the first quarter, triple the whole of 2025, and 95% of them AI-generated.
- A minute of generated video costs between $90 and $120, roughly 90% less than shooting the same minute with actors.
- Some performers are asked to hand their voice and likeness to a model before their contract ends.
Have an AI Sum Up This Article
ChatGPT95% of the quarter’s short dramas came out of a model
The number comes from the China Netcasting Services Association, the body that tracks online video output across the country. Roughly 128,000 short dramas were published in China during the first quarter of 2026, three times the total for all of 2025.
Of that volume, 95% came out of a generative model. The shift to AI video started with Seedance 2.0, the ByteDance video model that already rattled Hollywood, and then quality cleared another step.
Two releases finished the job of making human production optional on this format: Seedance 2.5 from ByteDance, and Wan 3.0 from Alibaba. The latter can now build thirty seconds of continuous video straight from a PDF.
The short drama format was almost designed for this substitution. Episodes of one to two minutes, repeating sets, standardised plots and scroll-based viewing: nothing in that brief ever required a performer to physically be somewhere.
Tripling output in a single quarter is really a statement about speed. An industry that took a decade to organise itself around studios, casting agencies and shooting calendars saw its volume multiply by three without any of those pieces needing to grow.
The size of the sector shows what is actually moving. Short drama production directly employed 690,000 people, and fifteen million Chinese workers list commercial livestreaming as their primary job. Both groups sit in front of the same wave.
Around a hundred dollars a minute instead of a shoot
Shen Yang, a professor at Tsinghua University, puts the saving at roughly 90%. A minute of AI video runs between $90 and $120, where the same minute shot with actors, a crew and a location sat an order of magnitude higher.
At that gap, the decision stops being a producer’s call. A studio that refuses generation ends up paying ten times more for content a competitor ships the same day, on the same platform, into the same recommendation slot.
That economics did not appear by accident: China now leads the global AI video ranking. On the Artificial Analysis evaluation platform, Wan 3.0 holds first place in text-to-video with audio, ahead of Gemini Omni, Google’s multimodal model.
Behind them sit a version of MiniMax H3 post-trained by the American platform fal, then the original open-weight H3, then Seedance 2.0. Chinese models occupy eight of the top ten places, and the same lead shows up in image-to-video and in editing.
The ranking method is worth noting, because it is not a lab measurement. Models are compared blind by users working from identical prompts, which pulls the verdict close to what a producer sees in real conditions.
Pricing follows the same curve as quality. The tariff war between Chinese labs, where Alibaba cut prices twelvefold on its fast model, applies to video too. Every cut makes a human shoot slightly harder to defend in front of whoever signs the budget.
More articles on Horizon
- Claude Code Limits Drop 17% on September 14
- Cursor Loses OpenAI Models on November 12
- Claude Sonnet 5 Fixes Claude Better Than 28 Researchers
When a studio distils your voice before letting you go
The harshest practice is not the dismissal itself but what comes before it. Some performers are invited to let a model learn their voice and their face, framed as a studio modernising its workflow, shortly before their contract quietly stops being renewed.
What is left is a job that still exists on screen and no longer exists on a payroll. Employment lawyers report a rise in AI-related labour disputes over the past two to three years, which places the start of this well before the current generation of models.
Nothing in that mechanism is specific to China or to entertainment. It needs three conditions only: work split into repeatable tasks, a digital record of past output, and a model good enough to reproduce the expected result without constant supervision.
Those three conditions turn up in voiceover, translation, dubbing, photo retouching, commercial visual production, moderation and a large share of customer support. Every one of those functions produces, daily, exactly the material used to train the thing that replaces it.
Timing matters more than diagnosis here. The rise in disputes over the past two to three years shows the substitution began with models far weaker than today’s, at a point when most of the professionals concerned still saw the subject as distant.
So the useful question is not whether a model can do your job, but whether your employer already holds the archive of your work. The answer is almost always yes, and there is a free report that measures how exposed your specific role is to that exact pattern. That archive, not the benchmark score, is what sets the calendar.
Follow the story on Horizon.
