What an AI Ad Actually Costs to Make in 2026

Get used to spending $100 to $200 on a good one and a half to two minute ad. That is the number I have settled on from my own projects, and it still holds in September. If you are not good at this yet, the same length ad can run you $500, and with Seedance 2.5 specifically, a poorly managed project can hit a thousand dollars before you notice.

That is a wide range for the same output, a finished ad of roughly the same length. The gap between the low end and the high end is not luck. It is almost entirely down to how much of your spend goes to generations you throw away.

A real ad, broken down

The clearest number I have comes from my own 1:45 cinematic ad, which I broke down credit by credit after finishing it. The piece cost about 115 euros to make, running on Dreamina at the time. That covered 26 video generations, a mix of 30, 10 and 5 second clips, mostly tested at 480p before anything went final, and 401 images along the way, references, sheets, tests, variations.

Four hundred and one images sounds like a lot for a minute and forty five seconds of finished video, and it is the part people underestimate when they budget a project. Character sheets, location sheets, object sheets, iterations on all three until they actually hold. Images are cheap individually, but the volume adds up, and skipping this phase to save money is exactly what makes the video phase more expensive, because a weak reference means more failed generations later.

Test cheap, spend once

The resolution you generate at changes cost, not behaviour. A generation at 480p follows the same physics and the same performance as the same prompt at 1080p, it is only the pixels that differ. So the discipline is simple. Test every shot at 480p until the performance, the timing, the camera move is right. Only then do you spend the credits to render at the resolution you actually need, 720p for social, 1080p for anything going to a client. Rendering finals at low resolution and hoping to upscale later is not the move either, upscaling has stopped being the reliable option now that native high resolution exists, so budget for one proper final render, not a cheap one you fix afterwards.

The habit that cuts spend by half

The single biggest lever on the whole budget has nothing to do with which platform or which resolution. It is reading the prompt Claude wrote before you spend a single credit on it. Claude hallucinates content into prompts constantly, a character with red shoes nobody asked for, an accessory that was never described, a camera move that drifts from what you actually wanted. If you do not catch that before generating, you pay for a generation you did not ask for, then pay again to fix it.

Reading every generated prompt line by line, out loud if it helps, before you run it cuts spend by fifty to seventy percent on my own numbers. It is tedious. It is also the difference between a $150 ad and a $500 one made from the exact same idea.

Where the platform and plan change the math

Seedance 2.5 is the most capable video model available and it is priced like it. If budget matters more than the last ten percent of quality on a given job, Kling 3.0 is the fallback, noticeably cheaper but a lot slower to work with, and it morphs on high motion and crowd shots in a way Seedance does not. Match the tool to the budget rather than defaulting to the most expensive option on every project.

Unlimited generation plans exist on some platforms and they are worth watching for, but read the fine print. Deals like this tend to be time boxed, around a month, and capped on generation length, so they suit a learning phase or a single intensive project far better than they suit ongoing client work. Discounts on the standard plans are real too. I have landed eighty five to ninety percent off around Black Friday, and there is a known trick of reaching the yearly payment screen on a platform, waiting, then clicking back, which has surfaced steep last chance discounts that were not shown up front. None of that is guaranteed to repeat, but it is worth checking before you pay full price.

Pricing it into the job

Know your real cost basis before you quote anything, and that basis is not one number, it is images plus tests plus finals plus the edit. A project that looks like it should cost $150 in generations can quietly become $400 once you count the reference images and the failed tests nobody budgets for. Build a small buffer into every quote for the generations that do not work, because some percentage never will, that is the nature of the model, not a sign you did something wrong.

What this post has not covered is what to actually charge a client for finished work, that is a separate question with its own answer, and how much AI creative work pays as a career is a different one again. This is only the spend side, what leaves your account to make the thing.

Common questions

How much does it cost to make an AI video ad in 2026?

My benchmark, as of September 2026, is $100 to $200 for a good one and a half to two minute ad. A poorly run project of the same length can run $500 to $1,000, mostly from generations spent on prompts that were not checked first.

How much of an ad's budget goes to images versus video?

More than most people expect. On my own 1:45 cinematic ad it was 401 images against 26 video generations for a total of about 115 euros. Sheets and references are cheap per item but the volume adds up, and skimping here makes the video phase cost more.

Does generating at a lower resolution actually save money?

Yes, and it does not change the output's behaviour. A generation at 480p performs identically to the same prompt at 1080p, only the resolution differs. Test at 480p, deliver at 720p for social or 1080p for client work.

What is the fastest way to lower AI ad production costs?

Read every prompt Claude writes before you spend a credit on it. Claude invents details that were never asked for, and catching that before you generate rather than after cuts spend by roughly half on my own figures.

Want the full method and the community that runs it every week? Join GenHQ.