Price the keep rate not the sticker
Effective cost is sticker times how many generates you throw away. Keep rate is the real budget.
A cheap model you reroll eight times is expensive. Effective cost is sticker price times how many generates you throw away. Keep rate is the real budget. Teams that only compare credit stickers end up surprised when the "affordable" path burns the monthly pool faster than the pricier model that lands usable takes on the first or second try.
This is not an argument for always buying the top tier. It is an argument for measuring the whole loop: prompt, generate, review, reject, adjust, generate again. The reject pile is where budgets hide.
How to calculate effective cost in credits
Start with a simple formula. Effective credits per keeper equals credits per generate divided by keep rate. If a generate costs 10 credits and you keep 1 in 5, you spent 50 credits per keeper. If another generate costs 18 credits and you keep 1 in 2, you spent 36 credits per keeper. The second path is cheaper in the only sense that matters: finished work leaving the studio.
Keep rate should be measured on your brief, not on a vendor demo. Demo prompts are flattering. Your brand pack, your product scale, your motion style, and your audio needs are the real test. Run a small bakeoff with the same reference kit and the same script. Count keepers with a harsh bar: would you ship this without apology?
Include partial keepers carefully. A take that needs heavy inpainting, lipsync repair, and frame fixes is not a full keeper. Either assign it a fractional keep or track a second metric: credits to ship-ready. Many teams discover that "almost good" outputs are the most expensive category because they invite endless repair.
Audio changes the math again. Silent plates plus a separate score can beat native audio paths when dialogue is not required. Native audio can win when lipsync and ambience must travel together. Price the path you actually use. Our breakdown in AI video cost per second with audio 2026 is a useful companion when you compare modes that look similar until you add sound.
Why sticker shopping misleads creative teams
Sticker shopping feels rational in a spreadsheet. It fails in production for three reasons.
First, human time is part of cost even when credits are the visible unit. Eight rejects mean eight review cycles. That is attention you could spend on hooks, captions, or distribution. Credits are not the only scarce resource.
Second, low keep rates create quality drift. Fatigue sets in. Editors start accepting weaker takes to stop the pain. Your channel's average drops while the spreadsheet still says you chose the "cheap" model.
Third, sticker comparisons ignore fixed setup costs. Building a solid reference stack, locking identity, and writing a stable prompt template takes effort. That effort amortizes better on a model that respects the stack. If a bargain model ignores your references half the time, setup work gets wasted every batch.
Use the credit calculator to sketch scenarios before you commit a campaign budget. Then validate with a timed pilot: one afternoon, one brief, two models, same reviewer. Write down keepers. Do not argue from memory.
Raising keep rate without romanticizing the model
Model choice matters. So do inputs. Soft references, conflicting style boards, and vague motion prompts crush keep rate on every system. Tighten the brief before you escalate spend.
Share one identity kit across a batch when you are testing hooks. Change one variable at a time. If you change face, wardrobe, and camera language together, you cannot tell what caused the rejects. Continuity discipline is a cost control tactic, not only an aesthetic preference.
Quote before you confirm large jobs. The habit in credit quote before you confirm prevents surprise drains when a longer duration or higher resolution multiplies the sticker. Know the credits at the moment of click.
Also separate exploration from production. Exploration can use cheaper, messier paths on purpose. Production should switch to the path with proven keep rate for that template. Mixing the two moods in one budget line is how "we only spent a little per generate" becomes "we spent everything."
A weekly budget ritual that stays sane
Once a week, pick your top shipping template (for example, 15-second UGC-style product hook). Log generates, keepers, and credits. Compute effective credits per keeper. Compare to last week. If the number spikes, look for process regressions before you blame the universe: new editor, new reference mess, new duration default, accidental resolution bump.
Set a soft ceiling for exploration credits so play does not eat the pool meant for client delivery. Versely costs are in credits, which makes this ritual straightforward if you actually record them. Guessing from vibes is how teams get blindsided mid-month.
If you want a clean place to run that ritual, sign up for Versely and keep quoting and batching in one workspace. The emotional tone worth aiming for is calm stewardship: spend where keep rate is real, cut where rerolls are a habit, and stop confusing sticker pride with finished work.
Common keep-rate killers you can fix this week
Several failures show up across teams. Mismatched first and last frames on product spots create morph artifacts that force regenerates. Conflicting style references (one cinematic still, one harsh UGC selfie) confuse the model and you pay for the confusion. Duration creep is quiet: someone bumps a five-second test to fifteen without updating the credit plan. Audio mode switches mid-batch make comparisons unfair because native sound takes burn differently than silent plates.
Fix these with checklists, not pep talks. Put duration, resolution, audio mode, and reference list in the project header. Reviewers should reject on a written bar: identity match, product readability, motion usefulness, and audio fitness. If your bar lives only in one senior editor's head, keep rate becomes a personality test.
Another killer is prompt churn without versioning. If nobody can paste the prompt that produced last Tuesday's keeper, you will spend credits rediscovering it. Store winning prompts next to the identity kit. Treat them like recipe cards.
Client feedback loops can also destroy keep rate when feedback arrives after you already burned the exploration budget. Align on references before the generate storm. A fifteen-minute reference review saves a hundred credits of polite disagreement rendered as video.
FAQ
Should I always pick the model with the best keep rate?
For production templates, yes, after you measure on your brief. For exploration, a cheaper messier path can be fine if you cap the credits. Match the path to the job.
What keep rate should I target?
It depends on the format. Many short UGC-style pipelines feel healthy when keep rate is high enough that reviewers are not exhausted. Track your own baseline for two weeks, then improve it. Absolute universal targets mislead.
Do failed generates still teach me something useful?
Sometimes. If you capture why they failed (identity drift, hands, motion, text, audio), rejects become training data for better prompts. If you ignore the why, rejects are only cost.