The sticker price of AI video is easy to look up — a few cents per second, less on a lighter model. The number on the invoice at the end of the month rarely resembles it.
The gap is retries. Published 2026 cost reporting puts finished video at roughly $13 to $220 per minute and names the dominant driver plainly: iteration rate, not per-second price. One survey found 68% of users exhaust their credits mid-project, because they budgeted for the shots they wanted rather than for the attempts it took to get them.
Where the attempts come from
Generation models carry no memory between shots. Describe the same person in the same words twice and you get two people who merely resemble each other. Practitioners testing character work across several tools reported needing 15 to 20 regenerations per scene to land something usable. Complex shots — fast motion, several characters, hands and faces — run a 30 to 40 percent regeneration rate on their own.
Drift is expensive in a way that is easy to miss. When a character changes partway through a sequence and you notice eight shots later, you are not repairing one shot. You are re-auditing everything downstream of the break.
Then there are the attempts that produce nothing at all. Kling users have reported 40 to 60 percent failure rates at peak hours, and vendor forums carry a steady stream of credits-gone-no-video reports. On most credit systems a failed generation draws down the same balance a successful one does.
What the field already recommends
The advice that recurs across cost guides is consistent enough to read as a checklist rather than as tips:
- Price the attempt before spending it. Know the shot count and the quote before the first billable call, not after.
- Resolve composition on something cheap. A still frame costs a fraction of a video render. Lock framing there before committing.
- Define the character once and reuse the reference. Re-describing a person in prose every time is what produces drift. A saved reference does not drift.
- Do not generate what the shot does not need. A shot whose start and end look nearly identical does not need two keyframes.
- Keep batches small. Six shots with a checkpoint beats twenty without one.
Making them defaults instead of discipline
Every item on that list is something a careful operator can do by hand. The difficulty is that all five have to hold on a Thursday afternoon, on the fourth revision, with a client waiting — which is exactly when they stop holding.
That is the part built into the Orkas video agent, so the list is enforced by the workflow rather than by memory:
- A quote before any billable call. The exact shot and character count, with a credit estimate, goes in front of you for approval. Nothing fans out silently.
- A default ceiling of six shots and three characters per video. Anything larger has to be proposed and approved first.
- A portrait anchor per character. Generated once, then locked and never regenerated. Later shots reference it by rule: the most recent frame from the same camera angle first, then an older frame, then the portrait alone.
- Keyframes spent by how much the shot changes. A near-static shot gets one keyframe, not two.
- Calls that never reached the provider are not charged. A local failure before dispatch consumes nothing.
None of this makes the first attempt land every time. No tool does that, and one that claims to is describing a demo. What it changes is the shape of the bill: the attempts you pay for are the ones you agreed to in advance.
What this does not cover
Cost control ends where publishing begins. Orkas carries a video from research and script through captions, thumbnail and search visibility, but posting it to each platform is still yours to do.