Sohu chief executive Zhang Chaoyang does not watch short dramas and watches AI-generated ones even less, as he has previously told Beijing News. At a media scrum during the Wuzhen World Internet Conference he added that the short-drama craze in China never reproduced in the United States and may be "a flash in the pan". [1] His timing coincides with the first full year of industrial-scale AI drama output, and the measurable part of the argument sits in a DataEye count. More than 221,900 AI dramas and AI comic dramas launched across Chinese platforms in the first half of 2026, in two formats, animation-style comic series and photorealistic series that imitate live actors; 1,055 of them passed 100 million cumulative plays, a 0.47 percent break-out rate, while over 90 percent sank into what the report calls cannon fodder. [2]
Generation capability has outrun the rest of the pipeline. ByteDance's Seedance 2.5 and Kuaishou's Kling 4.0 generate single takes up to 30 seconds and steer characters with reference inputs, with Kling 4.0 pushing 4K HDR output, by their makers' own published descriptions. [3] [4] What breaks batch production sits after the generator. Faces drift across long runs, a longer story has to be assembled from 30-second takes, and the surviving output then meets traffic buying (the paid ad pushes that place episodes in platform feeds) and payout coefficients the platforms set.
The plastic-feel remark predates Wuzhen. On August 6, after an offline event marking ten years of his English-class program, Zhang described AI-produced media as giving the world's media and content "a plastic feeling, like tomatoes forced by fertilizer, with no taste at all". [5] The remark resurfaced on November 9, when a Zhihu question page asking whether people agree had drawn more than 200,000 users into its thread. [5] At Wuzhen he also hedged. People love stories by nature, he said; viewers care about the plot and care little whether the performer is real, so whether short drama counts as an art form "remains to be proven by time". [1]
For anyone producing with these tools, the practical read is two checks. Judge consistency claims against episodes a platform has already released, since vendor demos do not carry through a long run. Treat the collapse in production quotes as a quality signal; when the going rate for low-end output falls from about RMB 5,000 per minute at the start of the year to RMB 200 or a few tens of RMB, the market is repricing footage viewers skim past. [6]
Seedance 2.5 and Kling 4.0 crossed the 30-second line
In July, ByteDance shipped Seedance 2.5. The company says the model generates up to 30 seconds in one pass, accepts as many as 50 multimodal reference assets, adds local editing inside a finished clip and natively handles more than ten languages. [3] Kuaishou's Kling 4.0 arrived in October, and Science and Technology Daily's launch report lists 10-bit HDR output at 4K in a 21:9 frame, 30-second native generation driven by an 8,000-token prompt budget, up to 15 reference assets spanning images, video, subjects and voice timbres, and 10 keyframes steering a continuation stretched to two minutes. [4] These are vendor-reported ceilings under controlled conditions.
The price of the generation step fell with the specs. CIC's managing director Zhang Chenkai told STAR Market Daily that a traditional top-tier short drama costs about RMB 1 million to produce, against RMB 80,000 to 100,000 for an AI photorealistic one, an 80 to 90 percent cut in total cost and up to 40 percent off the production cycle. [3] SeaArt's workshop claims master-tier Seedance 2.0 compute as low as RMB 0.3 per second, a vendor figure pitched at hundred-episode projects. [7]
ByteDance's own creation suite Jiemeng sells a character-lock feature whose stated job is fixing character face drift, which the vendor calls the biggest problem in comic-drama production. [7] A tool built to patch a defect is evidence the defect is routine.
Where a hundred-episode run stalls
CIC's research report describes the industry's own test as having moved from "can it generate" to "can it stably produce content that users consume and pay for". [3] The runs in question are long; vendors pitch the tools at hundred-episode projects, and one comic title, Zanjia Jianzong Tuanchong Xiaoshiwei, has run five seasons to 1,809 episodes. [7] [8] A long run needs consistency and narrative, and both degrade faster than image quality does.
The first fully AI-generated series to air in a satellite-TV prime-time slot, Hou Xiyouji, premiered on Hunan TV on August 31 and peaked at a 0.3384 percent rating with a 1.7862 percent audience share, first in its slot among provincial satellite channels. [3] Viewer complaints relayed in the same report describe the mechanics of the failure. The same surprised expression repeats across four or five shots, one viewer wrote, so that playing at double speed still feels slow. [3] Other viewers reported that different series generate faces so alike they blur together, with the verdict "there are too few quality titles; after two shows you are tired of them". [3]
A short-drama scriptwriter with two years in the trade told the same outlet that AI can draft a script in one click but the stories come out thin and stiff. "AI can assist, but someone still has to check and revise at the end," the scriptwriter said, adding that many scripts are now produced in volume with similar content. [3] Wang Qifan, founder of the script studio EverSail Script, put numbers on the squeeze. His live-action scripts hit at around 20 percent last year, and this year's rate has fallen as AI photorealistic supply floods the market "to chase volume or to chase the money". [3]
The money after generation
Tool fees are the small line item. Distribution and payout rules decide whether a finished series pays.
Production quotes collapsed across the year. Sina Tech's reporting traces the going rate for low-end dynamic comics from about RMB 5,000 per minute at the start of the year to RMB 200 per minute, and in some cases to a few tens of RMB, with many episodes running about one minute each. [6] Huang Haonan, founder of the studio Jiangyou Culture, quotes batch drama work at RMB 400 per minute, a total cost of RMB 40,000 or less for a hundred-minute title. [8] The three-tier market survived the collapse. Low-end dynamic comics, the cheapest comic-format tier sold by the minute, trade at tens of RMB per minute; mid-grade comic drama with human checking lists at RMB 1,000 to 2,000 per episode, off the start-of-year level; high-end 3D custom work keeps a premium that has narrowed. [6] Throughput scales with the price. Peng Xiang, who runs the studio Maopai Xingxing, says one company he has dealt with pushes out 600 episodes in a day at full tilt. [6]
The expensive line item sits after delivery. Wang Xukun, founder of the studio Xiaoxiang Yinghua, whose live-action team turns out 2,000 to 3,000 titles a year, puts traffic buying at roughly 60 percent of total cost in AI comic drama. [6] Industry accounts describe about RMB 70 of every RMB 100 in revenue going back into promotion. [8] Head teams have shifted from chasing hits to a "stable supply, protect the principal" approach; Wang says his first priority is getting his money back fast, and he estimates 80 percent of teams cannot get platforms to rate scripts before production starts. [6] Veteran producer Qi Yu calls the viral small-budget-big-profit stories "entirely survivorship bias"; most small teams' work sinks to the bottom of the content pool. [6]
Payout rules slid the other way. On the free-ad revenue-sharing models at Hongguo and Fanqie, the free short-drama apps where shows run ad-supported at no charge to viewers, settlement multiplies four inputs; monthly new effective viewing hours, a per-hour unit price, a content-type coefficient and a copyright coefficient. The unit price is not published and the coefficients are set by the platforms. [8] The trajectory has run one way so far. Around 2024 a title could still earn about RMB 100 per 10,000 plays; the figure fell to RMB 50, then RMB 30, and in 2026 many projects settle at RMB 5 to 10. One creator's backend shows 180 million plays settling at RMB 180,000, and some producers measure RMB 1 to 2 per 10,000 plays; at that rate a title costing RMB 20,000 to produce needs 100 to 200 million plays to cover production alone, before IP licensing and traffic buying. [8]
The cost mix inverted within the year. Accounts relayed by 36Kr put manpower, compute and miscellaneous costs at 6 to 3 to 1 early in 2026; manpower is now 30 to 40 percent and compute the largest item, with per-episode compute cost up from RMB 17 to RMB 80 to 100 after April price moves raised compute-card prices up to 34 percent, storage 30 percent, and some model input prices more than fourfold. [8] The compute figures include waste. One title kept 95 storyboard frames from 1,700 generated images, close to 18 discarded for every keeper, and each discarded frame was billed. [8] A solo creator's ledger lands the point in daily terms; about RMB 1 per second of compute including rejects and RMB 0.7 per second of editing, at most five minutes of finished footage a day, about RMB 510 in daily cost, against RMB 45 of revenue when a title reached 60,000 plays. [8]
A hypothetical example, labeled as such, shows the stack a newcomer would face. A 100-minute low-end run at RMB 200 per minute carries a production bill near RMB 20,000 at the batch rates above. Compute waste and promotion then multiply it; traffic buying takes roughly 60 percent of what a title spends in total, and the payout arithmetic above puts the same run in the 100-to-200-million-plays band needed to cover production alone. Every rate in this example comes from the reported figures; none of it is an offer or a forecast.
The rules themselves move mid-year. In February Douyin's short-drama copyright center set type coefficients of 60 for photorealistic drama, 50 for 3D and 40 for 2D, with commentary formats lower; in April it cut the photorealistic coefficient to 40 and slashed commentary formats by 80 percent; in July it lifted the commentary-comic coefficient from 1 back to 10, an adjustment the copyright center dated July 29, with bonuses starting at 50 million heat, the platform's popularity score for a title. [8] [7] Hongguo scrapped its series-scale coefficient from October. [8] Even with the tools, a watchable independent comic drama still takes three to five people seven to fourteen days, by the same relayed accounts. [8]
The outcomes follow the arithmetic. DataEye-ADX monitoring relayed by STAR Market Daily shows 98.7 percent of AI micro-dramas failing to recoup within six months of release, with 1.3 percent profitable, one in 77. [3] In the comic-drama traffic pool, hits with a return above one are under 0.1 percent, and roughly 200 titles are produced per break-out, by figures from the same DataEye report cited in a November aggregation crediting Guangming Online and Dawan News. [5] The number of active AI comic-drama contractors fell from 1,216 to 698 in the first quarter, by Qichacha statistics carried by Zhejiang Daily. [5] September's intake thins further; 51,800 new titles that month, about 82 percent of them under 1 million plays, while 1.46 percent of titles took 64.6 percent of the month's growth in plays. [8]
What the evidence settles
Zhang's quality complaint matches the record. His survival verdict sits outside it, and his own Wuzhen hedge, quoted above, already concedes the part the data cannot settle. [1]
The growth side of the market is real too. DataEye's report puts the January-to-May market above RMB 22 billion with a full-year estimate near RMB 40 billion, up 138 percent, and users past 600 million in the first half. [3] CIC projects RMB 85 billion by 2030 as launch volume climbs from 60,000 titles in 2025 to 170,000 in 2026. [3] Nomura's tracking puts Hongguo's daily active users at 168 million, up 107 percent, past the four long-video platforms combined. [3] Xinhua counts 430,000 micro-dramas launched in the first eight months, 13 times all of last year, with AI titles above 90 percent of them. [8] One relayed estimate puts this year's overseas AI drama market above US$4 billion, close to 70 percent of the overseas micro-drama pool. [5] Those growth figures and the 98.7 percent recoupment failure describe different sides of the same market; the revenue pool grows while the producers' economics do not. [3]
Break-outs exist and institutions are buying in. The AI comic Fapei Bianguan, Zuiqi Kaihuang Yangchu Zhanshen has passed 4.1 billion plays on Douyin across eight seasons, and Wanyao Tulu became the first AI drama on Hongguo to pass 100 million heat. [2] Bona Film's Sanxingdui Future Past became the first AI-made theatrical film cleared with a public screening license, set for October 23 at 100 minutes; the date was announced on September 27, and its shares opened limit-up the next day. [3] iQiyi's Liaoyuan program plans more than 40 AIGC films in 2026, with AIGC titles at 50 percent of newly green-lit projects. [3] DataEye's own report reads the pattern as IP-driven; series hits adapt from established web novels, while pure originals almost never become multi-season hits. [2]
Compliance is now a fixed layer. The National Radio and Television Administration's Order No. 16, the Measures for the Development of Micro-Dramas, passed on July 27, was published July 30 and took effect September 1. It defines a micro-drama as episodes shorter than 20 minutes, sorts titles into three filing classes by investment size and subject matter, and requires class-one and class-two titles to pass content review and obtain distribution permission before broadcast; unlicensed titles cannot air or enter awards. [9]
Chen Shaofeng, deputy director of Peking University's cultural industries research institute, told Sina Tech that AI short drama "has no future" as a content category, while AI short drama combined with brand communication is becoming a "big business". [6] On the record so far, the plastic-feel complaint lands where the evidence does, in sameness; the flash-in-the-pan call is a forecast about distribution and money, and nothing in the tool record settles it. The next readings of two numbers will move the argument further than another opinion will; the break-out rate off 0.47 percent, and the per-10,000-plays settlement off RMB 1 to 2.