I Made a 30 Second Commercial With AI for $45. Here Is Everything It Actually Took.

I just cloned myself. Twice.

Both men in the video below are me. I never turned on a camera, never booked a location, never hired anyone. The whole 30 seconds came out of one generation.

This AI video stuff is getting genuinely scary. But not for the reason people usually say.

What it cost

Everything ran through Seedance 2.5 on Higgsfield, priced in credits.

So the finished spot cost about ten dollars. Getting to a version worth finishing cost about forty five.

Worth saying that the reference set was free. Both character sheets and the location plate came off the free tier of ChatGPT, from one photo of my face. The only thing I paid for was the video generation.

The stills are worth calling out. A still costs 2 credits and a 30 second video costs 195. Three times tonight a 2 credit image caught a problem that would have cost 195 to discover in motion. That ratio is the single most useful thing I learned.

What it actually took

About three hours. Maybe ten minutes of that was prompting.

Here is the whole list, in order:

1. A script with a duration on every beat. Not "he pauses" but "a three quarter second beat." Every reaction timed, because timing has to be written down or the model invents it.
2. A stick figure storyboard of all nine shots. No faces, no likeness. It exists to catch composition problems before any expensive asset does.
3. Two character sheets built from one photo of me. Six angles, expression studies, a macro close up, and a full body against a height scale.
4. One location plate. A single frame carrying the whole room, which locks the colour grade so it cannot drift between shots.
5. A sixty second voice recording, trimmed to ten seconds for the reference.
6. A time check at 150 words per minute, comparing what each shot needs against the seconds it has.
7. Then, finally, the prompt.

That is basically the exact prep you would do before a real shoot. AI removed the shoot. It did not remove the filmmaking.

Three things that broke

The laptop pointed the wrong way for four renders. I had written the framing as "camera sees the back of the lid," which, worked through geometrically, points the screen at the wrong person. The glow on the other character's face was arriving from nowhere. Nobody noticed until it was pointed out, and then it was impossible to unsee.

The ending failed three separate times. A cut to empty landscape. A character who came back asleep. Then a cut that never happened at all. It only worked when I stopped asking for a crane and specified the drift the model kept defaulting to. Working with the tool beat fighting it.

The opening refused to cut. Four consecutive renders merged the first three shots into one long take, no matter how the prompt was phrased. The fix was deleting a shot, not rewording one.

What AI won, and what it lost

It won on cost, obviously. It won on speed, eight complete versions in one sitting. It won on consistency, the same face and the same colour grade across every shot, which is genuinely hard on a real set. And it did something no shoot could, which is put two of me in the same frame.

It lost on performance. The delivery is technically correct and completely empty. Every beat had to be specified in seconds, because a real person finds the beat. At one point I typed "he feels stupid for assuming it" into a prompt, which is a thing a human just does, in one take, without being asked.

It won every category I thought mattered, and lost the only one that actually does.

The part that actually matters

Making content is about to be basically free. Anyone can generate something that looks like this, in an afternoon, for the price of a decent lunch.

Which means the hard part stops being "can I make this" and becomes "is this even worth making."

More content is not the edge anymore. Knowing what to make is.

That is the whole reason Silma exists. Silma Research finds the formats that keep working in your niche, then turns the winners into ideas you can actually make. Not what looks like it is working because someone has a million followers, but what genuinely outperformed.

Because a thirty second ad you can generate for ten dollars is worth nothing if nobody wanted it.

Common questions

How much does an AI generated video actually cost? The final render of a 30 second 720p clip was 195 credits, roughly $10 on a $49 per month plan. Realistically budget four to five times that, because you will not get it right on the first attempt. Mine took eight renders.

How long does it take? About three hours for thirty seconds, and only around ten minutes of that was writing prompts. The rest is scripting, storyboarding, character references and timing.

Can AI video replace filming? It replaces the shoot, not the filmmaking. The prep is nearly identical. What it cannot currently do is perform, which is why anything depending on genuine emotion, comic timing or a real reaction still needs a person.

What tools were used? Seedance 2.5 via Higgsfield for the video generation, which is the only part I paid for. The two character sheets and the location plate were made on the free tier of ChatGPT with its built in image model. A few cheap test images came from Nano Banana on Higgsfield at 2 credits each. The end card was composited afterwards, not generated.