What an AI video ad really costs: our actual render bills

What does one second of AI video cost?

The video model we build on charges per generated second, and the price depends only on resolution. These are the list prices we pull from our provider's estimate endpoint, before any promotion:

Two things surprised me when I first saw this. First, price doubles with each resolution step, but on a phone the visible difference between 720p and 1080p in a Reels feed is small, because the platforms recompress everything anyway. Second, the voice and ambient sound are generated in the same pass as the picture. There is no separate text-to-speech bill and no editing bill, which is why a talking-head ad costs the same as a silent one.

We price everything at list, not at promo. During launch week our provider ran a 60% discount, and our first real 8-second 720p test ad cost about $0.32 instead of the $0.80 list price. It took about 4 minutes from approval to a finished file. We did not build our plans on the $0.32 number, because promotions end and a pricing page that depends on one is a pricing page that changes without warning.

Where does the money go in one finished ad?

Here is the full bill of materials for one default Addy ad: 20 seconds, Standard quality (rendered at 480p, upscaled to 720p), at list prices.

The brand brief is the step people expect to be expensive, because it feels like the smart part. It is not. After we moved our planning to a lighter language model on launch day, reading a site and writing the brief went from about 54 seconds to about 3 seconds, at roughly $0.002. Writing a batch of 8 concepts cost about $0.04 on the heavier model we started with, and we now budget $0.003 per ad for all the language-model work combined.

So the honest answer to "where does the money go" is: about 99% of it goes to rendering pixels. Everything clever that happens before the render is close to free. That has a practical consequence for anyone choosing a tool: the thing worth paying for is fewer wasted renders, not a smarter script.

Why does rendering at 480p and upscaling halve the cost?

Because the per-second price halves and the result still looks like normal phone footage. Our default quality, which we call Standard, renders at 480p for $0.05 a second. A finishing pass then scales it to 720x1280 with a Lanczos filter and applies a light unsharp mask, in the same ffmpeg encode that burns in the on-screen hook text. A 20-second ad drops from $2.00 to $1.00 of render cost.

Is it as good as native 720p? No. Fine textures like hair, fabric and small print look softer, and if you pause on a frame you can tell. In a feed, at scroll speed, on a phone screen, we have not found that it matters for testing. UGC-style ads are supposed to look like someone filmed them on a phone, and slightly soft footage fits that look better than a crisp, cinematic one.

Our rule: test hooks in Standard, then make the winners again in HD. HD uses twice the credits because it costs us twice as much. If you are deciding between resolutions for other reasons, our 720p vs 1080p guide goes deeper.

What does that mean for the price you pay?

Users never see dollars per second. They see credits, so plans can be compared with other UGC tools on the same scale. A 20-second Standard ad is 140 credits; HD is 280. Here is how that lands on each plan, straight from the numbers on our pricing page:

On Starter a 20-second Standard ad works out to about $1.35, and the shortest ad (15 seconds, Standard, 105 credits) is about 99 cents. Our cost for that 20-second ad is about $1.01, and Stripe takes 2.9% plus 30 cents of each payment. Credits are sized so every plan keeps a margin even if every credit is spent on HD at full list price. On Starter that margin is roughly 20%, which is thin on purpose.

Length matters more than anything else on your side of the bill. Because the render is billed per second, a 60-second ad costs four times a 15-second one. Ads up to 30 seconds are one continuous take; 45 and 60 seconds are three scenes with the same person, stitched together.

How can you cut your AI video ad cost?

These are the levers we have actually measured or built, in order of how much they save:

1. Do not render ideas you have not approved. Writing concepts costs fractions of a cent; rendering costs dollars. In Addy nothing renders until you swipe right on a concept. If your tool renders first and asks later, every rejected idea is a full render on your bill.
2. Render at a lower resolution and upscale. Half the cost for footage that holds up on phones. Save native HD for proven winners.
3. Keep ads short. Per-second billing means a 15-second test costs half a 30-second one. Find the hook in 15 or 20 seconds, extend later.
4. Write prompts that do not fail. Our biggest early waste was asking the model to draw things it cannot draw: on-screen text, logos, app screens. Those renders come back garbled and get thrown away. We now add text afterwards and describe products by shape and colour. More on that in our notes on what AI video models get wrong in UGC ads.
5. Make sure failures are refunded. Our provider refunds failed jobs, and we pass that through: credits are reserved when a job starts and returned automatically if it fails. Ask any tool you use what happens to your credits when a render fails.
6. Do not budget on promo prices. If a tool is cheap only because its provider is running a discount, expect the price or the plan limits to change.

What do these numbers not include?

A few honest limits on everything above.

• This is cost to make, not cost to run. Media spend on Meta or TikTok is separate and usually much bigger. A $1 ad still needs a test budget behind it; see how much to spend on creative testing.
• Our sample is small. Addy launched in October 2026. The per-second prices are list prices from our provider, and the timings come from our own test renders, not from thousands of production jobs. Render time in particular varies with queue load; we have seen about 4 minutes for an 8-second clip.
• Not every render is a keeper. Some clips come back with odd hands, a flubbed line or a product that looks wrong. We do not publish a reject rate because we do not have enough data yet to give you an honest one. Budget for some re-rolls.
• Model prices move. If our provider raises prices, our credit math changes and you will see it on the pricing page first. If they cut prices, we pass it on as more ads per plan.

If you want to see what a $1 ad looks like before you trust any of this, our examples page shows real outputs, and every format on the formats page lists its credit cost.

Is it cheaper to build your own AI video ad pipeline?

On raw cost per ad, yes, slightly: you would pay the same $0.05 to $0.20 per second we do, plus a few cents for a language model. That is the point of publishing this. The render is a commodity and you can buy it directly.

What you pay a tool for is everything around the render. In our case that means reading the brand, writing scripts that sound like a person, rules that stop the model from wasting renders, the upscale and hook-text pass, stitching long ads from scenes with the same face, refunds on failure, and posting to Reels, TikTok and Shorts. It took us weeks to get those right, and most of the cost savings in this post come from those rules, not from the price per second.

If you do want to build on top of us instead of from scratch, the Addy API exposes the same steps, and the MCP server lets Claude or ChatGPT call them for you.