How cheap did AI film-making get?

In the first half of 2026 the cost of a finished minute of AI short drama in China fell from about 5,000 yuan to a few hundred, and the country put out hundreds of thousands of micro-dramas, most of them AI-made. Supply ran far ahead of demand: of 221,900 new AI shows on Douyin in the first half of the year, 1,055 passed 100 million views. Our view is that this is a story about what becomes scarce, not about how cheap the tools got. When anyone can generate footage, the camera stops being the bottleneck and the audience becomes it, and a cheaper film is not the same as a better one.

The clearest figures come from CCTV, reported by Reuters on 25 September 2026: a finished minute of AI short drama cost about 5,000 yuan at the start of 2026 and only a few hundred yuan by mid-year. A film student in Hainan gave the change a shape rather than a rate card, describing a wedding scene that cost far less to make with AI than to shoot conventionally, helped further by local subsidies and a rent waiver. That is one person's example, not a published industry rate, and the true cost of a finished minute still depends on how many attempts each shot needs and how much human editing follows. The direction, though, is not in dispute, and it is why this sector moved so fast.

Local governments have treated AI video as an industry to be won, in the way China once approached electric vehicles and solar panels. Shanghai gives micro-drama producers computing capacity and hosted AI models; Beijing has set up a fund for film and video technology, and its Huairou district hands out vouchers that cut computing bills; Shenzhen offers AI film talent technical support and cheaper rent. A studio founder in Shenzhen told Reuters that officials from cities large and small now call him constantly to ask him to move his technology or his company. The platforms joined in too: iQIYI's chief executive said the company was going all in on AI and now subsidises some creators who publish AI content. We read the bidding as a sign of how badly the cities want the volume, not as evidence that the volume pays.

Volume is the part that should concern anyone planning to sell work. China's broadcasting regulator says hundreds of thousands of micro-dramas were released in the first eight months of 2026, most of them AI-generated, and that tens of thousands were removed for copyright, piracy or content violations. On Douyin, the data firm DataEye counted 221,900 new AI shows in the first six months and only 1,055 that passed 100 million views. Chinese rules already require clear AI labels, tightened for micro-dramas from 1 September, but copyright rules are still unsettled. The pushback is visible too: the box-office success of a crudely animated film made without AI was read by many as a reaction against the trend, and actors have complained about their likenesses being used without consent. The audience, in other words, is not a tap you can turn on by producing more.

There is no honest Singapore figure for AI production costs, and we will not borrow China's. What can be said from a local source is about demand rather than supply: The Straits Times reported on 11 August 2026 that microdrama apps made up half of the top 20 entertainment apps in Singapore by downloads and by revenue in May 2026, while the Chinese release pace kept climbing, almost all of it AI-made. Singaporeans are a market for this content, largely made elsewhere. The local production answer so far has been support rather than subsidy: IMDA launched a four-year Digital Content and Capability Development programme on 18 June 2026 to fund Singapore-centric content and experimentation with AI-assisted workflows.

For a small team, the useful part of the Chinese experience is not the volume, it is the splitting of the job. Reuters's reporting makes the point that a finished production still needs a coherent script, consistent characters, usable shots, editing and sound design, and that repeated generations quietly add to the real cost. Those are the same constraints we work under: the money saved moves from crew days to attempts, and an attempt that has to be made five times is not free. The genuinely new capability is narrower and more useful than the headlines suggest, scenes that were never affordable at all, made for the price of a prompt.

We would not treat the Chinese numbers as a target. They come from subsidies, a mature mobile-drama market and platforms that buy volume, and none of those conditions exist here. We read them instead as a description of the market's shape: when anyone can generate footage, footage stops being the scarce good, and finished, coherent, watchable work turns out to be rarer than the flood suggests. The figure in that Douyin data worth remembering is not 221,900; it is 1,055.

The discipline the cheap tools force is the part worth taking. Budget by shot rather than by day, decide which shots deserve the expensive model before you spend, lock your character and style references before any video credits go out, and judge the result by whether anyone watched it. Cheap generation has never made a story better. It has only made it cheaper to find out that the story was thin.