How to make AI videos: we made 3 reels and logged every dollar

Search for how to make AI videos and you mostly get lists of tools and their free plans. Almost nobody tells you what matters once you actually post: what a finished reel costs, how long it takes, and where the money goes.

On October 23, 2026, we made three reels end to end in one chat session with an AI agent. The agent made the stills, the clips, the voices, and the edit; we steered it with short messages. We logged every paid generation, every minute, and every redo. This guide is that log: the price of each step, what each reel cost, the method we used, and what went wrong.

These are the API prices we paid in October 2026, measured in the session, not copied from a pricing page. They change often, so treat them as an order of magnitude.

The takeaway is simple: video is almost all of the cost. A still costs about as much as a little over one second of clip. That's why the whole method below is about not generating video until you're sure of what you want.

The first reel cost four times as much as the third, and not because of a pricier model. It took about 19 stills and 5 clips to find the right scene. By the third, we knew what to ask for. The real cost of making AI videos is the price per attempt times the number of attempts, and attempts drop once you have a method.

1. Use AI for the hook only

All three reels had the same structure, 17 to 19 seconds in total:

• A 4 to 6 s hook: the only AI clip, with two lines of big text on top.
• A 6 s screen: the product doing the work.
• A 4.8 s scene: the result or payoff.
• An end card.

The screen, the payoff, and the end card are rendered HTML. They cost nothing and never look like AI. You spend the money where people decide whether to stay: the first seconds.

2. Still first, then the clip

Every clip waited for an approved still. If the still is wrong, the clip is wrong too, and the still costs 3 to 7 times less. On the first reel we made about 19 stills and only 5 clips. Once, the agent cancelled a queued $0.45 clip because we rejected its still, and we weren't charged.

3. Edit, don't regenerate

When something is off, ask for one change on the same still. We asked for "an Apple logo on the MacBook": one edit, $0.14, and nothing else in the picture changed. Regenerating from scratch changes the face, the light, and the framing, and you start over.

4. Cheap draft, good final

A 5 second draft on Seedance 2.0 Fast at 480p costs $0.29. The same final clip on Kling 3 Pro costs $0.56. Check motion and timing on the draft, and pay for quality once.

5. If someone talks, let the video model do the voice

For the car reel, the agent proposed two steps: a clip, then lip-sync, for $1.20. But Kling 3 Pro with audio says the exact words you put in the prompt, with matching lips, in one $1.01 step. The agent didn't suggest it; we found it by asking "the video model does the voice?". Always ask for the cheapest path that does the job.

6. Layer the audio per shot, and leave music to the app

In the car reel, her voice and a little road noise play only in the car shot; then the narrator takes over. Music isn't in the final file: trending sounds get added in Instagram or TikTok at posting time. Export with a clean voice track and note the second where the beat drops.

7. Check frames before you export

Before delivery, we checked frames at set times and found text covering a face (moved it), a word split across two lines, and one bad frame grab. Also check each platform's safe zones, where the app's buttons and captions cover the video.

This isn't a ranking of the best AI video generator. It's what we used in this session, for what, and why. We can only speak to what we paid for.

• Faces took more attempts than anything else. It took several rounds before a face landed. Sometimes the fix wasn't another generation but a closer shot: in the first version the person was small in the frame and her hand covered her face.
• AI tells slipped through. The first hook showed someone asleep under an absurd stack of paperwork. The reaction was "that's obviously AI". The fix was two plain, believable scenes, a kitchen and a car, with a normal amount of paper.
• The instant voice clone landed at "almost". From about 80 seconds of noisy phone clips, partly in Spanish, the clone "kind of sounds like her, but not fully". For a better clone: read a script for 1 to 2 minutes, in a quiet room, in one language.
• Waits are minutes. A Kling clip took 97 to 250 seconds, about 2 minutes typically, and lip-sync took 93 seconds. Meanwhile, the agent built the text, voices, and screens.
• Providers fail. The API service returned error pages several times and some uploads failed. Retrying fixed it and nothing was charged, but plan for it.

To try it out, yes: many tools have a free plan. But video models charge per second, and that cost gets paid somewhere: few credits, a watermark, lower resolution, or long queues. With our numbers you can estimate what a postable reel costs: $1 to $2 if you know what you want, and up to $5 if you're exploring. What saves the most isn't the free plan, it's fewer attempts.

• No real person's face or voice without their consent. The person in the third reel was there and agreed.
• No celebrity look-alikes. If someone asks for a famous person's look, it becomes plain features (hair, eyes, age, style) and an original person.
• An AI label on every video. Instagram and TikTok ask you to label realistic AI-made content.

An honest note: before this session, our reels on a new Instagram account got 3 to 10 views each, mostly from our own profile. Making the reel wasn't the problem; getting Instagram to show it was.

This workflow is the base of the new Hook Studio. It's a chat: you drop in images, video, or audio, and the agent builds the reel with the method above. It drafts cheap before the final, uses AI video only where it's needed, and shows you the cost in credits before every expensive step. At the end you get the video, the cover, and the caption with hashtags, ready to post.

If you're going to make AI videos, don't start by picking a tool. Start with the structure: a 4 to 6 second AI hook and the rest without AI. Approve the still before you pay for the clip, edit instead of regenerating, and always ask for the cheapest path. That took our cost per reel from $4.90 to $1.20 in a single session.