AI startup produces fully AI-generated feature-length film

Ronni Holmvig Strøm · 2026-08-22

Higgsfield AI has produced a 110-minute action-comedy called The Cully Hill Boys, generated end to end and made in four weeks for a reported $2 million. About half of it went to compute, to what the company calls AI tokens. A San Francisco startup did not make a feature film cheap. It moved the

AI startup produces fully AI-generated feature-length film

The Money Went Where the Compute Was

An indie feature usually takes a year or two and north of $20 million. The Cully Hill Boys cost roughly a tenth of that and shipped in a month, which is the part everyone will quote. But the interior of the budget is more instructive than its total. Roughly $1 million of the spend was pure inference, the cost of generating some thousand assets and a hundred locations on ByteDance's Seedance 2.5 model. The other million covered the licensing, the screenplay, the direction, the finishing. The things people do.

That first million is the number to watch, because it sits on a curve that has only ever pointed down. Every model generation makes the same output cheaper to produce. Reprice the compute half at the next Seedance, or the one after, and the film that cost $2 million costs $1.25 million, then less. The human half does not fall on that curve. The human half is the floor the price settles toward.

The Human Layer Held Its Value

The most convincing part of the film was the writing, and the writing was done by a person. Higgsfield hired a human screenwriter, Tim Planagan, and the authorial voice underneath the images is what holds the thing together. Semafor's reviewer noted that the dialogue does not read as machine-written. The images, by contrast, sometimes announce themselves as generated, and the weaker scenes read as what one critic called prompt fatigue, the AI equivalent of bad acting.

So the pipeline did not collapse into a single prompt box. It reorganized. The generative model absorbed the parts of filmmaking that were expensive because they were labor-intensive: sets, locations, coverage, the sheer volume of footage. What survived as valuable was judgment. Someone still had to write a story worth watching, direct performances built from licensed likenesses of Israel Adesanya and Rampage Jackson, and decide which of a thousand generated assets was good enough to keep. Forty percent of the crew had never used these tools before the project began. They learned them in four weeks, which tells you the barrier was never the craft of prompting. It was knowing what to make.

Open-Sourcing the Playbook Is the Real Release

Higgsfield published the prompts, the production logs, and the workflow documentation for the entire production, the same thing it did with its earlier $500,000 feature Hell Grind. A studio guards its pipeline. A platform hands it out, because the platform's interest is in ten thousand people trying the workflow next, not in one film's box office.

This is where the optimism is earned rather than asserted. A licensing deal for real celebrity likenesses and a documented pipeline anyone can copy, backed by a $1 million contest to pull creators in. That is not a machine replacing filmmakers. It is the cost of attempting a feature falling far enough that the people who could never raise $20 million get to try, and the part that stays hard, the part where you need something to say, becomes the part that decides whether the result is any good.

The open question is what happens the next time the compute half halves. When tokens cost a quarter of what they cost this August, the human layer will not be the smaller line on the budget anymore. It will be almost the whole of it.