What an AI brand film costs — the published numbers

The cheapest AI brand film on the public record cost about two thousand dollars. The most ambitious one involved around a hundred people. Both ran as national advertising. That gap is not a measurement error and it is not vendors lying. It is the one variable that decides what this work costs: whether there is a character in it who has to stay the same person from shot to shot.

What follows is everything we could find publicly on the record. We have not published a teardown of our own budgets, so none of these numbers are ours. Almost all of them came from the party that made the work, which is the first thing to hold in mind while reading any of them.

What is actually on the record

The low end is a single spot. A thirty-second commercial for the prediction market Kalshi ran during Game 3 of the 2025 NBA Finals. It was made by one person, the solo AI filmmaker PJ Accetturo, using Google Veo 3, with Gemini writing the shot list and ChatGPT used for script ideas. The reported figures are around two thousand dollars, three days from idea to air, and three to four hundred generations to yield fifteen usable clips. It cleared network approval. Source is independent reporting rather than a studio release, which makes it the most useful number here and also the most abused.

Made in

3 days

Crew

One

Generations

300–400

The Toys"R"Us origin film, premiered at Cannes in June 2024, was made by Native Foreign with Toys"R"Us Studios using Sora before public access. About a dozen people, roughly three months on and off. NBC News noted that this was comparable to a normal job. No saving was claimed and none is visible.

Coca-Cola's 2025 holiday films, produced with Silverside AI and Secret Level, involved around a hundred people, five AI specialists over about thirty days, and more than seventy thousand generated video clips for two films. Secret Level founder Jason Zada has described the crew comparison as roughly twenty people against fifty or more on a traditional shoot. That is a vendor claim and should be read as one. Seventy thousand clips is the number that tells you something the crew figure does not.

Svedka's Super Bowl LX spot on 8 February 2026 was the first primarily AI-generated national Super Bowl ad, made by Silverside AI for the Sazerac brand. It took about four months, mostly to reconstruct the Fembot character and train the model. Humans wrote the storyline. CMO Sara Saunders called it a way to start a conversation, and was explicit that it was not a cost or time saving. A brand volunteering that its flagship AI production saved nothing is worth more than any case study.

McDonald's Netherlands released an AI Christmas film in December 2025 and pulled it after backlash. Agency TBWA\Neboko, production company The Sweetshop. Sweetshop CEO Melanie Bridge said ten people and five weeks, in a LinkedIn post that has since been deleted. Vendor claim, no longer standing behind itself, and the film is not on air.

Where a real saving has been reported

One agency figure holds up better than the rest. The US independent Copacino Fujikado told Digiday it delivered a campaign requiring twenty-six models across several locations for twenty-six per cent of live-shoot cost. Still an agency describing its own job, but specific and structurally believable. The brief was logistics. Most of what a conventional production would have spent went on moving people to places, and that is the part generation actually removes.

What the brands claim

Unilever's digital-twin product imagery figures come from a press release issued with NVIDIA and Collective World: twice as fast and fifty per cent cheaper overall, an eighty-seven per cent cost reduction on TRESemmé Thailand with a five per cent lift in purchase intent, and up to fifty-five per cent savings with sixty-five per cent faster turnaround and double the click-through in Beauty and Wellbeing. That is a brand and vendor announcement, unaudited, and it is about product stills rather than film.

“A brand volunteering that its flagship AI production saved nothing is worth more than any case study.”

Mondelez told Reuters, through Jon Halvorson, that it has invested over forty million dollars in a generative AI system built with Publicis Groupe and Accenture, targeting a thirty to fifty per cent cut in content production costs, with short TV spots aimed at the 2026 holiday season. Source is interview-based reporting from October 2025. We found no evidence that those spots have run. Until they do, the only verified number in that story is the forty million spent, not the saving promised.

Cost is also moving the other way

The saving narrative has a counterweight that rarely gets quoted. Forrester data cited by Fortune puts the growth in the cost of agency business at eighty-three per cent in 2025, with only seven per cent of agencies selling generative AI capability as a separate service. Tool cost is largely being absorbed rather than billed. Someone is paying for the seventy thousand clips, and at the moment it is often not the client.

Nobody has a rate card

Digiday surveyed how agencies actually bill compute in March 2026 and the answers do not agree with each other. Merge passes token costs through to clients, metered. Big Spaceship treats compute as a production cost, in the same bucket as catering or equipment hire. Pencil, part of Brandtech, sells generation credits priced against chat responses, images or seconds of video, with volume commitments. Silverside bills a per-seat subscription. Horizon Media describes its fee as cost recovery, not revenue. Lerma/ shows tokens at cost as a line item. RPA and Kepler absorb it entirely. No agency disclosed a markup. No published rate card exists anywhere in this category.

Big Spaceship's answer is the one that reads as settled thinking. Compute is a production cost. It behaves like one, it scales with ambition like one, and pretending it is a software licence hides it from the people approving the budget.

The comparison that does not exist

There is no like-for-like named budget comparison in the public record. Same brief, AI against conventional, both dollar figures reported, same brand, same deliverable. It has not been published. Every percentage you have seen quoted is built on something less than that.

Which makes the most-repeated claim in the category traceable to one source. Anyone saying AI production is ninety-five per cent cheaper is extrapolating from Kalshi, and Kalshi was one operator, one thirty-second spot, no talent, no legal review of a likeness, and no brand-safety committee. Scale that to a brand with a character, a music clearance and six markets and the arithmetic stops being arithmetic.

So the only reliable way to price this work today is to price the thing that drives it. Ask how many characters the film has, how many of them have to be recognisably the same person in every shot, and how many generations the studio expects to run. Then ask which line on the quote is compute, and whether it is at cost or marked up. Nobody in the Digiday survey would answer that last question publicly. A studio should be able to answer it for you in a meeting.