AI video credits: the real cost of paying per generation

Credit-based AI video tools are cheap to start and hard to budget, because you pay for attempts, not results. A finished ten-second shot might take one generation or fifteen, and the meter runs on every one of them. That makes the real cost of a project unknowable until it is over, which is the opposite of how producers want to work.

This article explains where credits actually go, gives you a simple way to estimate a project before you start, lists the questions to ask any credit-based vendor, and looks at the alternative: AI that runs on your own machine and is included in the plan you already pay.

How credits really get spent

Pricing pages show a cost per generation or per second. Real projects spend credits in places the pricing page does not emphasise.

• Iteration. The first output is rarely the one you keep. Wrong framing, a strange hand, a face that changes between shots. Each fix is a new generation.
• Variations. Directors want options. Four variations of a shot cost four times as much, even if only one is used.
• Resolution and length. Higher resolutions and longer clips usually cost more per attempt, so the expensive settings are also the ones you iterate on most.
• Upscaling and extensions. Extending a clip, upscaling it or changing its aspect ratio are often separate charges.
• Failed jobs. Depending on the vendor, a generation that errors out or is blocked by a content filter may still consume credits. Check the policy rather than assuming.

The result is a cost structure where the more care you take, the more you pay. That is a tax on iteration, and iteration is where quality comes from.

Estimate a project before you start

You cannot remove the uncertainty, but you can bound it. Use a simple attempts-per-keeper ratio.

1. Count the shots that need generation and their length in seconds.
2. Run a small test. Generate three or four representative shots and count how many attempts each needed before you had a usable one. That gives you a realistic ratio, perhaps four or six attempts per keeper.
3. Add variation and revision rounds. If the client will see two options per shot and you expect one round of notes, multiply accordingly.
4. Price the extras. Upscales, extensions and reframes for other aspect ratios.
5. Add a margin. A 30 to 50 percent buffer is reasonable on a first project with a new tool.

Write the estimate down and track actual spend against it. After two projects you will know your real ratio, which is worth far more than any pricing page.

Questions to ask any credit-based vendor

Before committing a production to a credit system, get clear answers to these:

• Do unused credits roll over, and do they expire?
• Are failed or filtered generations refunded?
• What does each resolution, duration and model tier cost?
• Are upscaling, extending and reframing billed separately?
• What happens to the price when a new model version is released?
• Can generated material be used commercially, and under what terms?
• Can I export clean, high-quality files and separate elements, or only a compressed MP4?

The last question matters as much as the price. A cheap generation that arrives as a single flattened file can become expensive when every note means generating again. We explore that in editable project vs locked MP4.

Where paying per generation still makes sense

Cloud generation is the right tool for some jobs. When you need footage that does not exist and cannot be shot, such as an impossible camera move or a concept visual for a pitch, paying for generations can be far cheaper than a shoot. Large cloud models can also do things a laptop cannot.

The problem is not credits as such. It is paying per attempt for tasks that are routine, repeated across every project and perfectly capable of running locally: transcription, subtitles, translation drafts, voice-over drafts, upscaling archive footage, reframing for social. When those everyday tasks sit behind a meter, the meter becomes a reason not to try things.

On-device AI included in the plan

The alternative model is AI that runs on your own computer. You pay once for the plan, the processing happens on your hardware, and the cost of the tenth attempt is the same as the first: some of your Mac's time.

cinemaOS works this way. The AI features run locally on Apple silicon and are included in cinemaOS Complete, €9.99 or $9.99 per month excluding VAT, with no credit meter.

• Transcription with Parakeet or Whisper, with speaker detection for up to four speakers.
• Local translation of subtitles into 58 languages, with honest quality tiers. Dutch, French, German and Portuguese (Portugal) are rated Good; other languages are Draft or Rough and need review before delivery. Our guide to translating subtitles locally on a Mac explains the workflow.
• Dubbing with Quick Voice-over or Production Dub using macOS voices, and optional voice cloning only after a voice-rights attestation.
• Restoration and upscaling with bundled models, plus Reframe with subject tracking and tracked Face & Object cover.
• J.A.R.E.D., the on-device assistant, for edits on the timeline.

Nothing is uploaded for AI. No project, media, transcript or prompt leaves the Mac, which also matters for client confidentiality. See on-device AI video editing and privacy.

The trade-offs are real. Local processing uses your Mac's time and memory; on an 8 GB machine, heavy AI jobs run one at a time. Local models are smaller than the largest cloud models. And cinemaOS is not a text-to-video generator, so if your project needs footage invented from a prompt, you will still use a cloud tool for that part. The sensible setup for many teams is both: pay per generation for the few shots that truly need it, and run the routine work locally.

FAQ

Are credits always more expensive than a flat plan?

No. For a single small job with a handful of generations, credits can be cheaper. The difference shows over many projects, and on tasks you repeat constantly.

How do I stop a team from burning credits?

Set a per-project budget, test prompts at low resolution first, and agree in advance how many variations a client will see. Track spend per shot, not just per month.

Does local AI need a powerful Mac?

cinemaOS needs Apple silicon and macOS 15 or later, with 40 GB of free disk recommended. 8 GB of memory is supported; heavier jobs simply run one at a time.

Next step

Run the attempts-per-keeper test on your current tool and see what your real cost per shot is. Then compare it with work that does not need a meter. cinemaOS includes its on-device AI in the plan, and the 60-day free trial needs no card.