WORKFLOW • Framework updated · Jul 2026
How to Budget Prompt Iterations for AI Video
Turning a first prompt into a usable AI video clip almost always takes several attempts, and most platforms bill credits for every generation whether or not you keep the result. Without a record of what each attempt changed and cost, iteration spend drifts upward invisibly. This guide gives you a simple tracker template and a set of stop-rules to keep that spend predictable.
Why every attempt is billable
Most AI video tools charge per generation, not per kept clip, so a throwaway test still draws down your credit balance. On Runway's developer API pricing, Gen-4 Turbo is 5 credits per second and Gen-4.5 is 12 credits per second, so a five-second Gen-4.5 test costs 60 credits regardless of the outcome. Luma bills per generation and rounds duration up, so a nine-second request is billed as ten seconds. Some models also enforce a floor: Runway's Aleph 2.0 is 28 credits per second with a 56-credit minimum per generation, which makes many tiny retries costlier than they look. Because the meter runs on attempts rather than on keepers, the only reliable way to control cost is to track each attempt deliberately and decide in advance how many you can afford.
The four-field iteration tracker
Track every generation as one row with four fields. Prompt version is a short label or number, so you can tell attempt three from attempt seven without rereading the whole prompt. Change made records the single thing you altered since the previous version, such as adding a lighting cue or one background element. Credit cost is the exact credits that attempt consumed, calculated from the model and duration before you run it. Verdict is a one-word judgment: keep, discard, or revisit. Luma's own guidance supports this structure, advising creators to note the keywords that consistently produce results they like and reuse them, which is exactly what the change and verdict fields capture over time. Luma boards also retain context and remember earlier generations, giving you a built-in place to keep versions side by side rather than restarting from scratch each time.
Estimate credit cost before you click generate
Fill in the credit-cost field before running, not after, so a bad run is a decision rather than a surprise. Runway sells API credits at $0.01 each in its developer portal, so an attempt's dollar cost there is credits per second times seconds times $0.01; a ten-second Gen-4.5 attempt is 120 credits, or $1.20. Luma's per-generation charges vary sharply by tier: Ray 2 at 720p is 160 credits for five seconds and 320 for ten, while Ray 2 Flash at the same resolution is 55 credits for five seconds and 110 for ten. Pika's model 2.5 charges by resolution and duration for text and image-to-video: a five-second clip is 24 credits at 480p, 20 at 720p, and 40 at 1080p on paid plans, and ten-second clips cost more. Stamp each figure with the model version, resolution, and duration, because the same prompt on a different tier bills differently.
Stop-rules that cap runaway spend
Three rules keep a promising idea from eating a month of credits. First, run early exploration on the cheapest tier available: Luma publishes a Ray 3.14 Draft Mode at 20 credits for five seconds and 40 for ten, roughly an eighth of full Ray 2 720p cost, so burn early prompt versions there and move up only once the composition is right. Second, set a hard attempt cap against your plan's monthly ceiling before you start. Pika's Standard plan gives 700 credits per month, which is about 17 five-second attempts at 1080p, so if a single clip is not working by attempt eight, stop and rethink the prompt rather than the model. Third, watch for per-generation minimums like Runway's 56-credit floor, which quietly make short retries expensive. Write your cap into the tracker as a line you will not cross.
Keep the tracker current as models change
AI video pricing and model lineups shift often, so a tracker is only useful if its cost figures are dated. Runway lists a hard sunset of July 30th, 2026 for Gen-3 Alpha Turbo and Gen-4 Aleph, recommending Gen-4.5 or Gen-4 Turbo in place of the former and Aleph 2.0 in place of the latter, which means a rate you logged against a retired model no longer applies. Record the date you last confirmed each rate, and reverify on the provider's official pricing page before a large batch. For heavy, repeated iteration, check whether a flat option fits: Luma's Relaxed Mode, available on its Unlimited and Enterprise plans, runs at lower priority but lets you create generations without topping up, which changes the math when you expect many attempts. Match the plan to how much you iterate, not just to a single project.
Editorial note: This framework is general information, not a vendor endorsement. Check the current pricing, terms, and data-handling details directly with the provider before buying.