What a second of AI video costs
Video generation services charge per second of finished clip. Models are easiest to compare on one price list, so we use Runway's developer pricing page: it lists models from several makers, including Google, Kling, Wan and Seedance, in credits of $0.01. Below are five models whose price is given specifically for 1080p, from cheapest to most expensive.
The National Bank of Ukraine rate on 20 October 2026 is UAH 45.0013 per dollar. The clip is 15 seconds, a typical length for feeds and short video.
| Model (1080p) | Price per second | One 15 s generation |
|---|---|---|
| grok_imagine_1_5_lite | $0.14 | $2.10 (UAH 95) |
| gemini_omni_flash_1.1 | $0.15 | $2.25 (UAH 101) |
| wan3 | $0.20 | $3.00 (UAH 135) |
| seedance2 | $0.40 | $6.00 (UAH 270) |
| seedance2_5 | $0.68 | $10.20 (UAH 459) |
The same models can be bought directly from their makers. For example, on the Gemini API pricing page Google bills gemini-omni-1.1-flash by tokens: 5,792 tokens per second of 720p video at $17.50 per million, roughly $0.10 per second. That page gives no separate 1080p price. Runway also has a ready product_ad recipe for a product ad clip: 216 credits for the first 4 seconds at 1080p and 40 for each further second, which is $6.56 for 15 seconds.
A generation budget: why attempts matter more than the model's price
One generation rarely gives a clip you can run. The model may spoil hands, a logo or the product's shape, and then the generation is repeated. So the budget is:
clip seconds × price per second × number of attempts × number of variants
An ad test needs several variants, not one; how to split a test budget between them is covered in the article on an advertising budget for an online shop. Take six 15-second variants and five attempts each. Five attempts is a conditional figure: yours depends on how complex the scene is, and it is worth measuring on the first clip.
Calculation: 15 s × price per second × 5 attempts × 6 variants × 45.0013. 1080p prices from Runway's price list and the NBU rate on 20.10.2026; the number of attempts and variants is conditional.
Two practical conclusions follow.
- Models differ by multiples, and so do attempts. A cheap model that gives a usable clip on the ninth attempt costs as much as an expensive one on the third. Compare models by the price of a usable clip, not of a second.
- Generation is not the only cost item. There is also the script, selecting attempts, editing, voice-over and captions. These costs depend on your team, so we do not estimate them, but they belong in the estimate.
We do not compare these sums with a shoot: there is no public, verifiable price list for filming an ad clip. Compare it yourself: ask a videographer for a quote on the same six variants and put it next to the line in the chart plus your own person's editing hours.
When you still need a shoot
The decision is set not by the budget but by what the clip claims. Ukraine's Law "On Advertising" (Art. 1, item 11) calls advertising unfair when it misleads or may mislead through inaccuracy, unreliability, exaggeration or omission, and Art. 10 prohibits such advertising. A generated frame is not a violation in itself. The violation is when it shows something untrue about the product.
Hence three cases when the clip is better filmed:
- The clip demonstrates a product property. The fabric does not crease, the paint covers in one coat, the vacuum picks pet hair off a carpet. A model shows what you described in the prompt, not what the product can do. If a property is the ad's main message, it has to be filmed on the real product.
- The clip shows a specific product bought for its looks. Colour, texture, size in the hand. A model reproduces the product approximately, and the buyer sees the difference after delivery. Product photos in Merchant Center have separate requirements; see the article on AI product photos.
- A person in the clip talks about their experience. A review nobody gave, a generated "customer", "doctor" or "expert" is an invented person confirming your claim. The risk here is no longer just the label but the content of the ad itself.
Generated video is safe where it claims nothing verifiable: background and atmosphere, logo or infographic animation, abstract transitions, an illustration of an idea without a promised result.
AI labels in Google and Meta
Both platforms require a mandatory disclosure for ads about elections, politics and social issues. For ordinary commercial ads the rules differ.
✓ — the rule is stated unambiguously; ≈ — depends on the condition in the cell; ✗ — not possible. Meta's help page describes images in ads; it has no separate rules for video. Sources: Google Ads help "Use AI content label settings and disclosures" and Meta help "How AI-generated images in ads are identified and labeled on Meta" on 20.10.2026.
What follows for a Ukrainian business:
- Ads in the EU. If a Google Ads campaign serves in EU countries and you labelled the asset as generated, the label appears on the ad itself. Plan the creative so the label does not cover an important part of the frame. Google warns separately that its setting does not guarantee compliance with local laws.
- Your own label in the frame. Google allows you to add your own "made with AI" label directly into a clip or image and does not treat it as breaking the rules on text overlays. Do not put it in the very corner: the platform may crop the edges for different formats.
- Metadata. Google embeds an invisible SynthID watermark and C2PA metadata in clips made with its tools. Meta says it labels images from third-party AI services too. So if a clip is generated, do not count on it "going unnoticed".
Models disappear faster than campaigns
On 20 October 2026 the Gemini API pricing page carries a warning: three preview versions of Veo 3.1 shut down on 22 October 2026, and Google advises moving to gemini-omni-1.1-flash. That is why we did not put those models' prices in the table.
For a business this means one thing: do not build a process around a specific model name. Keep prompts and settings in a form that can move to another model, and check the price list before each new campaign, not once a year. The price per second can change, and the model that gave you a good style can disappear.
Generate or film
- Does the clip show what the product can do?A property the buyer will check after purchase. Yes → film it on the real product
- Does the buyer choose by this product's looks?Colour, texture, size. Yes → the product is filmed; background and transitions may be generated
- Does a person in the frame talk about their experience?A review, an expert, a "customer". Yes → only a real person, with their consent
- Is the ad about elections, politics or social issues?Here the label is not your choice. Yes → disclose it in the campaign settings
- Does the campaign serve in the EU?A labelled asset in Google gets a label on the ad. Yes → leave room for the label in the frame
A first step for a week
- Pick one clip from your next campaign that claims nothing about the product's properties: a mood, a background, a sale announcement.
- Generate it on two models from the table, a cheap one and a dearer one, with the same prompt.
- Count every attempt. Note on which attempt each model gave a usable clip.
- Multiply the attempts by the generation price: that is your real price per clip for each model.
- Label the asset as generated in the Google Ads settings and run it next to an ordinary creative.
Our position
Generated video is worth using for ad variants that convey a mood or an idea and claim nothing the buyer will check after purchase. The generation price is too low to be a reason to refuse. What limits it is something else: what the clip shows about the product and how the platform labels it. Labelling a generated clip yourself, rather than waiting for the platform to do it, is cheaper than explaining to buyers why the product looks different.
The position does not hold if the ad's main message is a property or the look of a specific product: then it has to be filmed, and generation stays for background and transitions. Nor does it hold for ads on political and social issues, where disclosure is mandatory and an error has other consequences, or for products whose advertising is regulated separately (medicines, medical services, financial services): there you need a lawyer first. The budget calculation assumes five attempts per clip; measure your own figure on the first clip and recalculate.
Verified on 20 October 2026 on Runway's developer pricing page, the Gemini API pricing page, the Google Ads help on AI labels, Google's political content policy, Meta's help on AI labels in ads, Ukraine's Law "On Advertising" and the National Bank of Ukraine rate. The arithmetic was recalculated in code. The number of attempts and variants is conditional. This article is not legal advice.