The sticker price on an AI video model is a lie of omission. The tool says fifty cents a clip, and you do the math on the clips you need. Then you actually generate, and the bill is three times what you planned, because you paid for every attempt that did not make the cut.

That is the whole trap in 2026. A per-attempt pricing model charges you directly for uncertainty. No model lands a usable clip every time, so the real cost is never the price of the keeper, it is the price of the keeper plus the four rerolls it took to get there.

There is one number that fixes how you think about this, and almost nobody tracks it: your keeper rate. This is the playbook for raising it.

The real number is cost per usable second

Stop pricing your work by the advertised per-clip rate. Price it by what a usable second actually costs you.

Keeper rate is the share of your generations you actually ship. If you generate ten clips and keep three, your keeper rate is thirty percent, which means your effective cost is not the raw credit price, it is roughly three times that. Across the industry that multiplier lands somewhere between 1.7x and 3x for most people, and higher when prompts are vague or the shot is hard.

So the cheapest model is not the one with the lowest sticker. It is the one that gets you to a keeper in the fewest tries on the shot you actually need. Optimize the keeper rate and the bill takes care of itself.

Step 1, measure your keeper rate before you change anything

You cannot improve a number you are not looking at. For one real project, count two things: how many generations you ran, and how many you actually used. That ratio is your baseline. Most people are shocked .. they assume they are at sixty percent and they are at twenty-five.

Do this per shot type, not just overall. Talking heads, fast motion, hands on a product, and crowd scenes have wildly different keeper rates, and the average hides which one is eating your budget.

Expected result: a baseline keeper rate for the project and a clear read on which shot type is burning the most credits per keeper.

Step 2, spend the cheap decisions before the expensive ones

Every choice you can make in text is free. Every choice you discover by generating costs a credit. So front-load the free ones.

Lock the framing, the one camera move, the lighting direction, the wardrobe, and the action in writing before you generate anything. Vague prompts do not fail cheaply, they fail at full price and then you reroll blind, changing three things at once and learning nothing. Decide on paper, generate to confirm, not to explore.

Expected result: a fully specified shot brief, so your first generation is a test of one thing, not a coin flip on five.

Step 3, prove the shot on the cheap model first

Not every shot needs your premium model. A talking head at a calm motion level is something a mid-tier model handles fine, and the price gap between tiers is large .. the premium models run several times the cost of the value ones.

Use the cheapest model that can plausibly do the shot to confirm your framing, your timing, and your composition. Only spend the premium credits once the plan is proven, on the final pass where the extra quality actually shows. Proving a layout on an expensive model is paying steak prices to taste-test.

Expected result: your composition and timing confirmed on a low-cost model, with premium credits reserved for the final render of shots that need them.

Step 4, the reroll-versus-repair rule

This is the single biggest credit saver, so internalize it. Before you hit generate again, ask whether the problem is local or structural.

Local .. one bad frame, a flickering edge, a small artifact in an otherwise good clip. Repair it in post. Do not reroll a good shot to fix a one-second blemish.
Structural .. the motion collapses, the hand is wrong across most frames, the composition is off. Reroll, because no amount of post fixes a shot the model never got right.

Half of all wasted credits come from rerolling a near-keeper that a thirty-second repair would have saved, and the other half come from endlessly post-processing a clip that was structurally broken from frame one.

Expected result: a habit of classifying every reject as local or structural, repairing the locals, and only rerolling the structurals.

Step 5, swap the model instead of grinding the same one

Some failures are a model signature, not bad luck. A given video model may simply never nail a specific hand pose, a particular fast motion, or a certain camera move, and you can feel it when reroll three looks exactly like reroll one.

When two rerolls fail the same way, stop. That is the model telling you this shot is not its strength. Swap that one shot to a different model rather than spending ten more credits forcing the first to do something it cannot. The fastest way to blow a budget is loyalty to one model on the shot it is worst at.

Expected result: a two-strike rule .. after two same-failure rerolls, the shot moves to a different model instead of a third reroll.

Step 6, batch the QC pass

Reviewing one clip at a time, regenerating, then reviewing the next is how you lose an afternoon and a wallet. Generate your batch, then watch every clip once at full speed in one sitting, marking keep, repair, or reroll.

Batching does two things. It makes the reroll-versus-repair call faster because you are comparing clips side by side, and it stops the emotional reroll, the one where you rerun a perfectly fine clip because you stared at it too long and convinced yourself it was off.

Expected result: one full-speed QC pass over the whole batch with each clip tagged keep, repair, or reroll, instead of a clip-by-clip rerun spiral.

Where this runs

The reason rerolls spiral is usually friction. If swapping a problem shot to a different model means re-uploading your references into a different tool, you will grind the first model instead, because the swap is annoying. If your cheap-model test and your premium final live in different apps, you will skip the cheap test.

A single canvas removes that friction, which is what makes this playbook actually cheap to follow. On Vilva every major image and video model is on one board, so the Step 5 model swap is one click with the same references already attached, the Step 3 cheap-to-premium escalation is a model dropdown not a tool migration, and the agent keeps your locked brief and assets attached across every reroll so you are testing one variable, never five. The keeper rate climbs because the cheap decisions stay cheap. Free to try at vilva.ai (200 credits on signup).

Troubleshooting and next steps

My keeper rate is fine but the bill is still high. You are over-specifying premium shots. Move more confirmation passes to the cheap model (Step 3) and reserve premium for finals only.
I keep rerolling the same clip. You skipped Step 5. Two same-failure rerolls means swap the model, not a third try.
Repairs take longer than rerolls. Then it was structural, not local. Re-read Step 4 .. if most frames are wrong, reroll; only repair truly local blemishes.
The cheap-model test never matches the premium output. That is fine. You are testing composition and timing, not final quality. Lock the plan cheap, render it premium.
Next step: track keeper rate per shot type for a month. The one shot type that stays low is the one to template .. a fixed prompt, model, and motion level you reuse instead of rediscovering every project.

The takeaway

AI video does not get expensive because the clips cost too much. It gets expensive because you pay full price for every clip you throw away.

Watch the keeper rate, spend the free decisions first, repair the locals and reroll the structurals, and swap the model before you grind it. The bill is just the keeper rate wearing a disguise.