What does an AI video really cost after retries?

It is your total generation and retry spend divided by the number of output seconds that passed an acceptance rule you wrote down before generating. Formula: (generation + retry spend) ÷ accepted usable seconds. It differs from advertised per-clip prices because it counts every failed take, and it is undefined — not zero — until at least one second passes acceptance.

Record every attempt for each shot: what you were charged (including failed takes), whether the result passed your pre-agreed acceptance rule, and how many usable seconds it produced. Then divide total spend by total accepted seconds. Only seconds that passed the rule you froze before generating go into the denominator; shortening the goal afterwards to make the number look better invalidates the comparison.

A credit is a platform-internal unit: each platform prices models, resolutions and durations differently in its own credits, so 100 credits on one service is not 100 on another. The comparable figure is cash spent per accepted usable second of output for your specific task, which is what this worksheet measures.

A second of output that passes the acceptance rule you defined before generating — typically covering subject or product fidelity, motion, camera behaviour, continuity, artifact tolerance, duration, resolution, aspect ratio, format, and independently verified commercial-rights status. Passing is defined by your written rule, not by whether the clip "looks good".

Four levers, in order: pick a model matched to the task rather than the cheapest sticker price; test risky elements with small cheap runs before committing to the expensive final take; set a stop budget per shot before you start; and re-use configurations that already passed acceptance instead of re-discovering them. Lowering the acceptance bar is not a cost reduction — it moves the cost into rework.

For each shot: the model and settings used, whether it passed the frozen acceptance rule, whether it passed on the first attempt, total attempts, accepted usable seconds, everything you were actually charged including failures, time to the first accepted result, and the main failure reason (subject mismatch, motion/camera, continuity, artifact, format, rights). Ten shots is enough to expose a pattern without turning the test itself into a budget.

There is no single figure, because the cost of a generation is the model multiplied by the settings you choose. Across the models routed on this site the baseline cost runs from 40 credits (Gemini Omni 1.1 Flash) to 570 credits (Google Veo 2). And the sticker price of one generation is not your real cost — the real cost includes every retry it took to get a result that passes your acceptance bar, which is what the calculator above measures.

Because each table quietly fixes one configuration and compares that. Change the resolution, the duration, or whether audio is on, and the ranking between tools reorders. A price per video is only meaningful next to the configuration it assumes, which is why every row in the snapshot table on this page states its configuration.

No. When a job reaches a failed state on this site the credits charged for it are returned to your balance automatically. On platforms where failures are not refunded, those charges belong in the spend column of the worksheet above — they are part of your real cost.

It is free, requires no login or email, and everything you type stays in your browser. Nothing is uploaded: no prompts, no media, no worksheet text. A draft is kept in your own browser storage so you can come back to it, and you can clear it with the Reset button.

Run the idea at the lowest viable setting first — shorter, lower resolution, no audio — and only re-run the take that worked at the quality you actually need. Most of your discarded attempts should be the cheap ones. Then track the test in the worksheet so you know your real cost per usable second, not just the per-run price.

No. They are a snapshot taken on 2026-09-08 using the same calculation the workbench performs, and providers change their rates. Treat the table as a planning aid and the figure shown in the workbench before you generate as authoritative.