Cutting AI generation costs by a third without touching quality

Generative video is priced per second, and the difference between providers is large enough that the naive approach of picking one model and using it for everything quietly costs multiples of what the work requires.

We were paying $0.84 for a five-second clip. We now pay $0.56 for the same clip at the same quality. The change was not a cheaper model.

The default that costs money

The model generated synchronized audio by default. We were producing clips for a pipeline that replaces the audio track downstream anyway. So we paid a premium, every render, for a soundtrack that was discarded seconds later.

Turning it off cut a third off the bill and changed nothing about the output we actually used.

Why a registry beats a hardcoded default

The obvious fix is to pass the flag. The durable fix is to stop hardcoding provider quirks in application code, because every provider names things differently and prices them differently. We moved model metadata into a registry: which parameter controls audio, whether it affects price, what the per-second cost is. Job code asks for what it wants and the registry resolves how to ask for it.

That structure also makes an unknown model a loud error rather than a silent wrong default, which matters more than it sounds. The expensive failures in cost engineering are the quiet ones.

The habit worth stealing

Read the pricing page for every parameter, not just the model name. Then ask, for each thing you're paying for, whether it reaches the final output. Ours didn't, for months, because nobody had checked what the default was doing.

Resolution defaults compound downstream

A second default cost us more than money. Asking an image model for a "16:9" frame produced dimensions that were not exactly 16:9. Close enough to pass a glance, wrong enough that every downstream step inherited the error.

Feeding that image into a video model produced a clip at yet another near-miss aspect ratio, and by the time it reached delivery the output was neither a standard size nor a clean crop away from one.

The fix was a free local step: a small utility that crops or pads to exact target dimensions and normalizes the pixel aspect ratio, run at the end of every job. Generative models are approximate about geometry; standard image tooling is exact. Let each do what it's good at.

One caution learned the slow way: a still image and a video frame need different handling. Forcing a video-oriented pixel format onto a still will flatten transparency and shift colour. The utility has to know what kind of file it's fixing.

Where to look first in your own pipeline

• Every parameter that is on by default. Ask whether its output reaches the final deliverable. Ours didn't.
• Every dimension you request. Verify what you actually received rather than what you asked for.
• Every model chosen once and never revisited. Generative pricing moves constantly, and last quarter's best value frequently isn't.
• Every retry. Failed generations bill the same as successful ones, so an unreliable step is a recurring cost, not just an annoyance.

Inference cost is the operating expense that scales exactly with your success. It deserves the same scrutiny as any other line item, and it usually gets none.