Late on a Thursday night, Yeh stared at one red cell in her spreadsheet.
Yeh is an illustrative character in this article, not a real customer. She is an independent producer with a short film project of about 10 minutes that she plans to generate entirely with AI. Her pricing logic is simple: a live-action shoot means paying a crew, renting locations and renting gear, and AI does not, so the total should come down a lot.
In the first column of her spreadsheet she had written "how much AI saves." The red cell was in the last row: she could not work out a total price.
The First Thing the Panel Overturned Was the Idea of Saving Money
The panel was a streamed session hosted by InVideo, a video tool company, with directors Rajkumar Hirani and Shakun Batra and several other panelists. When reading the optimism about AI tools in the discussion, it is worth factoring in the host's position.
Shakun Batra said in the panel that he does not see AI as just a cost-optimization tool. People often ask whether a film that costs 2 crore rupees could be made for 30 lakh. His view is that once you start playing that game you have already lost, because it is the least interesting goal, and what is worth doing now is imagining more. He also said that as a writer, he always carries a floor of production risk in his head while writing, this cannot be written and that need not be written, and he hopes to find writers who are not bound by that line.
Yeh paused when she read this. Her entire spreadsheet was built on exactly that game.
Hirani's 10 Minutes Showed Her Which Cost to Look at First
Rajkumar Hirani talked in the panel about his experience making a centenary film for Bajaj. When the brand approached him, his biggest worry was that he could not find an actor who looked close to Gandhi, and there was also makeup, shooting locations, and family members who would have to appear on camera. He decided to try AI, partly as a way to learn.
The approach went like this. The team shot still photos of the family members from many angles in Pune and took voice samples, with another studio handling the quality of the voice. Then they had other actors perform for a day and a half and transferred those performances onto the faces. The finished film ran about 10 minutes.
What really made Yeh stop was another part of his account. AI teams are used to pursuing one perfect shot at a time, but what he wanted was "generate it first and give it to me, and I will cut it." He gave an example: suppose a story has a hundred shots, with someone generating the second shot and someone else generating another, and in the end everything gets mixed together and nobody knows which shot is where. Along the way faces change, and teeth sometimes grow bigger than any normal person's. He said the two sides argued about this many times.
None of these arguments appears as a line item on any quote.
She Got Her First Number and Became Even Less Willing to Quote
Yeh next read the budget discussion in the middle of the panel. One panelist said he had worked out the per-minute cost of his previous production and then used that figure to estimate a full feature, only to find he could not afford it, and that live action might even be cheaper. That was a film with built sets and a period setting.
Another panelist said his team had made detailed budgets for four films, and that they keep building a baseline for compute cost. He asked himself whether taking 1,500 to 2,000 dollars as the standard is reasonable, whether it should be higher or lower, and whether more iterations would be needed, and admitted there is no answer yet. The panel did not make clear what unit that figure is measured in, and it is not a conclusion, so this article does not treat it as a market rate, and readers should not apply it directly. He added something more worth noting: revenue has not changed, the theaters, streaming platforms and music labels are still the same group, and buyers also know that input costs have fallen, so the price they are willing to pay is pushed down too.
As for where cost actually increases, someone on the panel was quite specific: first, the number of iterations, and second, the gap in understanding between creators and the technical team. Someone on the panel also mentioned that for some shots, getting one usable take might mean burning through a hundred or more first.
Yeh picked up a pen and wrote a formula of her own beside the spreadsheet. The following is an illustrative calculation, and the numbers are purely hypothetical:
Total generation cost ≈ number of shots × price per generation × average number of retries
Suppose she has 100 shots, the price per generation is counted as 1 unit, and each shot takes 5 tries on average. The total is 500 units, not 100. If each shot averages 8 tries, it is 800. The unit price did not change, yet the bill is 60 percent higher.
Subsidized Prices Will Expire, but the Cost of Redoing Will Not
Another part of the panel pushed that formula one step further for Yeh. One person said platforms will raise subscription fees later, and another picked up on this to say that roughly 70 percent of current prices is subsidy. That is a panelist's verbal estimate, with no source or method of calculation given in the panel, so this article reports it as stated and does not treat it as fact.
She therefore did two things. First, she split the quote into two separate columns, "unit price" and "retry count," and estimated them separately instead of folding them into one total. Second, she left one row at the bottom of the budget for a "unit price increase" scenario.
The same panel also gave her a method. Someone expects that very soon no studio will give a green light without a proof of concept, and a single "show me two minutes first" could replace ten rounds of persuasion. Someone else said that AI's greatest value is lowering "the cost of being wrong." What that panelist meant is this: the success rate in the film business is low, so what really slows growth is that failure is too expensive, and being able to redo things cheaply is healthier than simply making things cheaply.
So Yeh made "make two minutes first" the very first step of the whole budget.
On Friday Morning, She Had One New Habit
The next morning she changed the red cell in her spreadsheet into a range, and noted next to it that this was a hypothetical scenario and the numbers would be filled in after real testing.
She also first checked how each of the models she was using is priced, how much credit was left in her account, and what the last test round had actually cost. She used to treat these as after-the-fact reconciliation work. Now they were things she needed to see before shooting began.
This is also the natural place for a gateway like AI Token King. According to the service's MCP documentation, the same API key can be used to query available models and pricing, the account balance, and a paginated billing usage record. Image and video generation go through the same interface, and every actual generation call deducts account credit. The documentation also explains that checking the balance once before and once after a call lets you work out the deduction for that call. This means that "how much did the retries cost" is a number that can be reconciled and recorded.
The next time someone says in a meeting that "AI makes it cheap," Yeh will probably ask first: for one shot, how many tries?
FAQ
Is AI filmmaking cheaper than live action?
The views in the panel were not uniform. One panelist worked out a per-minute cost, used it to estimate a full feature, and found it unaffordable, with live action possibly cheaper, especially for a film that needs built sets and a period setting. Another panelist argued that setting the goal as making the same film for less money is itself the least interesting game. Whether it pays off depends on the type of film and the number of retries.
Where does the cost of AI video mostly go?
Two items were named in the panel: the number of iterations, and the gap in understanding between creators and the technical team. Someone on the panel also mentioned that for some shots you might need to generate a hundred or more before picking one usable one.
How do I estimate AI filmmaking cost?
You can start by splitting the quote into two columns, "unit price" and "average retries," and then multiplying by the number of shots. For example, the illustrative formula in this article is total generation cost ≈ number of shots × price per generation × average number of retries. The numbers should be filled in from your own test shoot, and the numbers in this article are only hypothetical.
Will AI generation prices stay this low?
One panelist said that roughly 70 percent of current prices is subsidy and that platforms may raise subscription fees later. This is a verbal estimate with no source attached and should not be treated as fact. A safer approach is to leave a row for a "unit price increase" scenario in your budget.
How can I tell what each retry cost?
If you call models through the AI Token King gateway, its MCP documentation says you can query available models and pricing, the account balance, and a paginated billing usage record, and you can reconcile using the difference between the balance before and after a call. Which models are actually available depends on the platform's model list.
Source Disclosure and Editorial Notes
This article is based mainly on a panel video hosted by InVideo, with its content reorganized and context added, and it is not a sentence-by-sentence translation. The panel video does not label who is speaking for each remark, and the identity of some panelists cannot be confirmed from the on-screen text, so wherever the attribution of a statement could not be determined, it is written as "another panelist." Yeh is an illustrative character, not a real customer, and the formula and numbers in the article are hypothetical scenarios, not an actual quote. The cost-related remarks in the video are the panelists' verbal opinions, have not been independently verified, and do not constitute budgeting or investment advice.
Video title: What AI Filmmaking Really Costs ft. Rajkumar Hirani & Shakun Batra. Channel: Invideo for Filmmaking (this panel was hosted by InVideo). Original link: https://www.youtube.com/watch?v=nOiIa8vqgGc
Further reading on aitoken.com.tw: how-to-build-ai-video-generator-reduce-cost-seedance, ai-token-real-cost-vs-subscription-price, and why-same-ai-coding-task-costs-so-differently-across-tools.
Turn Every Retry into a Number
Before your next quote, it may help to run one test round first, write down how many times each shot was actually retried, and then go back and correct those two columns. To look up pricing for each model, you can refer to the model and pricing information from AI Token King at aitokenking.com.tw/home.
