Is AI video generation a waste of money in 2026?

Whether AI video generation is a waste of money needs a fresh answer in November 2026. The market has changed quickly. OpenAI announced in March 2026 that it would close the Sora app in April and the API in September,¹ and the Sora 2 API was shut down on 24 September 2026 without a named replacement.² Google shut down the Veo 3.1 Preview in the Gemini API on 22 October 2026 and now points developers to Gemini Omni Flash, or to the generally available Veo models on its enterprise platform.³ A price list more than a few months old is probably wrong.

This post sets out what the main tools cost today, why they cost what they do, the costs the price list leaves out, the labelling rules, and where paying for AI video makes sense for a business. How video is found in Google, on YouTube and in AI answers is covered in the image and video SEO guide.

Most tools now charge by the second of generated video, and the price rises with resolution and with sound. The table uses list prices from each company's developer platform, because that is where per-second prices are published. Google Cloud's price list shows Veo per unit of output; Google has quoted Veo by the second since at least September 2025, when it priced Veo 3 at $0.40 a second.⁴ Consumer apps sell the same models through monthly credit plans.

The cheapest tiers now cost a few cents a second, and the spread is wide: a second of 4K from Seedance 2.0 costs about 30 times a second of 480p from MiniMax H3-Lite. Runway also sells several other companies' models through its own credits, at prices close to the makers' own: Veo 3.1 with audio costs 40 credits a second there, the same $0.40 as Google's list price.¹⁰ Kling, Seedance and MiniMax are all made by Chinese companies, the subject of the guide to Chinese AI models.

Consumer plans hide the per-second price inside credits. Runway's Standard plan costs $15 a month for 625 credits, which Runway says buys 52 seconds of Gen-4.5 or 104 seconds of Gen-4 Turbo.¹³ Kling's Standard plan costs $6.99 for the first month and renews at $8.80.⁹ Work out the seconds a plan buys before comparing plans on their monthly price.

A video is a long run of images that have to agree with each other, and the bill follows the pixels. BytePlus publishes the formula it uses for Seedance: input plus output duration, times width, times height, times frame rate, divided by 1,024.¹² By that formula a five-second 1080p clip at 24 frames a second comes to about 243,000 tokens, which is why the same clip costs $1.87 at Seedance 2.0's 1080p rate. Gemini Omni Flash is also billed in tokens, at 5,792 output tokens for each second of 720p and 17,376 for each second of 4K, three times as many.⁵

The computing behind those tokens is heavy. Julien Delavande, Régis Pierrard and Sasha Luccioni measured open text-to-video models on an Nvidia H100 GPU in a 2025 paper and found that cost grows quadratically with the spatial and temporal size of the clip, so doubling the resolution or the length more than doubles the work. One short clip from WAN2.1 1.3B used about 90 Wh, which they put at roughly 30 times an image generation and 2,000 times a text generation.¹⁴

That cost lands on the providers first. When OpenAI ended Sora in March 2026, a spokesperson linked the decision to growing compute demand, and Engadget reported Appfigures data showing installs and spending on the app falling at the start of the year.¹⁵ Where providers compete, prices are still falling. Google charged $0.75 a second for Veo 3 until September 2025,⁴ and its cheapest Veo tier now costs $0.05 a second with audio at 720p.⁵

The per-second price buys a second of generated footage. A business needs a second of footage it can actually use. The gap between the two is where most of the money goes.

Retries you pay for

Every take is billed, including the ones you throw away. BytePlus charges only for successful generations, so a clip blocked by moderation costs nothing, but a clip that renders and is simply wrong is billed in full.¹² Veo 3.1 can return up to four videos from one prompt,⁶ and per-second billing means you pay for all four. In a controlled test by Ipsos and Syracuse University's Newhouse School, some of the AI ads needed multiple iterations.¹⁶

Clips too short to use on their own

Veo 3.1 makes clips of 4, 6 or 8 seconds,⁶ and Kling 3.0 makes clips of 3 to 15 seconds.⁹ That suits a cutaway in a social post. A 30-second ad or a two-minute explainer has to be built from several clips, each generated, checked and paid for separately.

Keeping people and products consistent between shots

Each generation starts afresh, so a face, a product or a logo can change from one shot to the next. The vendors sell fixes for this. Google's January 2026 update to Veo 3.1 promised better identity consistency for characters, settings and objects across scenes, using reference images.¹⁷ Those features help, but every attempt to make shot four match shot one is another billed take. For a product that has to look exactly like the one in the box, this is both a cost and a risk.

Editing and finishing

A folder of clips still needs someone to cut it together, add the logo, captions, music and offer, and check every claim it makes. That is a person's time, and no per-second price includes it.

What a 30-second ad costs to generate

A worked example shows the scale of the generation bill. An invented florist wants a 30-second ad made of five 6-second shots at 1080p with sound, and keeps one take in four. The keep rate is my assumption for the sums; I know of no study that measures a typical one. That means generating 120 seconds of video to publish 30.

Even the dearest option costs under $50 in generation. What decides whether the money was well spent is the time spent prompting, judging takes and editing, and whether the finished ad performs as well as one made another way.

Rights and commercial use

Check the terms of your plan before a clip goes into an ad. Kling's free Basic plan says generated content is not for commercial use; its paid plans allow it.⁹ Runway does not claim ownership of outputs or restrict their commercial use, but its terms say your use of outputs is at your sole risk, disclaim any warranty of non-infringement and give Runway a licence to use what you generate.¹⁸ Google lists its enterprise Agent Platform API, used with generally available Veo models, among the services covered by its generative AI indemnity,¹⁹ and its terms limit that cover to use that is not provided free of charge.²⁰

Ownership is less certain than the terms suggest. The US Copyright Office concluded in January 2025 that AI output can be protected only where a human author has determined sufficient expressive elements, and that prompts alone do not count, though a human's creative arrangement or modification of the output can be protected.²¹ A raw generated clip may belong to nobody, which could leave you little to stop a competitor reusing it in the US. The position differs between countries, so take advice for the markets you sell in.

Platform risk

Tools close. Sora users had to export their work before the deadlines, after which their data was due to be permanently deleted.¹ Google's documentation gives 17 November 2026 or later as the retirement date for Veo 3.1 and Veo 3.1 Fast on its enterprise platform.⁶ Buying a year of credits upfront is a bet that the tool, the model and the price will all still be there.

YouTube

YouTube requires creators to disclose realistic content made or altered with AI that makes a real person appear to say or do something they did not, alters footage of a real event or place, or shows a realistic scene that did not happen. Clearly unrealistic content, beauty filters, colour correction and AI help with a script do not need a label. YouTube may add the label itself for content made with its own AI tools, content carrying C2PA metadata or content its systems detect, and creators cannot switch it off in the first two cases. Disclosure alone does not limit a video's audience or its eligibility to earn money, while creators who keep failing to disclose risk having labels applied for them, removal or suspension from the Partner Program.²²

Google's tools can mark footage at the source. Veo 3.1 on Google Cloud supports Content Credentials (C2PA),⁶ and Google says every Gemini Omni video carries an invisible SynthID watermark.²³ If that metadata survives your editing, YouTube may label the video automatically.

The EU AI Act

Article 50 of the AI Act has applied since 2 August 2026. Its main duty on businesses that use AI video tools, which the Act calls deployers, covers deep fakes: AI image, audio or video that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Such content must be disclosed clearly, at first exposure at the latest, and the watermark a tool maker embeds does not meet the duty on its own. Fines can reach €15 million or 3% of worldwide turnover.²⁴

For evidently artistic, creative, satirical or fictional work, the duty is limited to disclosure that does not hamper the display or enjoyment of the work. The tool makers carry a separate duty to mark their output in a machine-readable format.²⁵

UK advertising rules

The CAP and BCAP Codes have no AI-specific rules, and the existing rules apply however an ad was made. The ASA suggests asking whether the audience would be misled without a disclosure, and says a label is very unlikely to fix a fundamentally misleading message. Its example is an AI-generated image of a cosmetic's effect that does not reflect real-world results.²⁶

What audiences think of AI ads

Audiences are cooler on AI ads than advertisers think, but labelling an ad did not hurt it in one test. In IAB's survey of 505 US Gen Z and Millennial consumers and 104 ad executives, run from October 2025 to January 2026, 82% of executives believed young consumers felt positive about AI-generated ads. Only 45% of the consumers did, and 39% of Gen Z felt negative.²⁷ A MediaScience study with the Ehrenberg-Bass Institute, reported in May 2026, tested four kinds of AI label on 900 US respondents and found no significant difference from an unlabelled control in brand choice, brand attitude or ad liking. MediaScience has also developed an AI video platform, which is worth bearing in mind.²⁸ In the Ipsos test, 79% of consumers said companies using AI should have to disclose it.¹⁶

• The free routes. Google made Gemini Omni available at no cost in YouTube Shorts and the YouTube Create app,²³ and since September 2026 anyone with a Google account can generate 1080p video in Google Vids at no cost.²⁹ For Shorts and internal videos, the cheapest option is often free.
• Short social clips and b-roll. Clips of 3 to 15 seconds match a cutaway, a background loop or a product moment in a Reel, which is the job current tools suit.
• Product-led ads with a clear message. In the Ipsos and Newhouse test of 20 ads from 10 brands, shown to 3,000 US consumers, the AI versions did well on product-driven briefs with a clear problem, solution and reason to believe.¹⁶
• Testing an idea before a shoot. A rough AI version of a concept is a cheap way to see whether a script works before paying for a crew. I would treat it as a sketch for the brief.
• Animating your own photography. Veo generates from a reference image as well as a text prompt,⁵ so existing product photos can become short motion clips without a new shoot.
• Volume at the cheap tiers. At $0.05 a second, an 8-second Veo 3.1 Lite clip at 720p with audio costs $0.40 to generate.⁵

• Brand storytelling. In the Ipsos test, human-made ads were on average 14% stronger on short-term effectiveness and 17% stronger on long-term brand equity, and the AI versions struggled with briefs that needed emotion, a creative leap or a point of view.¹⁶ Those AI ads were made by Gemini and Sora 2 with no human creative input, so a skilled team using AI as one tool might close the gap. The study did not test that.
• Showing results that are not real. An AI depiction of results a product does not deliver is likely to mislead, and the ASA says a label is very unlikely to fix that.²⁶ Realistic footage of real people or events also triggers YouTube's disclosure rules and, in the EU, the deep fake duty.
• Long videos with the same person or product throughout. With clips capped at 8 to 15 seconds and consistency costing retries, a two-minute explainer with a recurring presenter is an expensive thing to generate shot by shot.
• Mass-produced YouTube channels. YouTube's monetisation policy, renamed in July 2025 to cover inauthentic content, makes AI-generated content built on generic or unoriginal templates that give the impression of mass production ineligible to earn money.³⁰
• 4K that nobody needs. Veo 3.1 costs 50% more a second at 4K than at 1080p,⁵ and Seedance 2.0 about twice as much.¹² Pay for 4K only when the placement uses it.
• Large credit bundles bought upfront. Kling's first-month prices renew higher, and tools can close within months, as Sora did.

Neither Google nor YouTube documents any ranking benefit or penalty for how a clip was made. The ordinary rules apply.

• No video result without a watch page. Google indexes videos on pages where watching one video is the main reason to visit, and gives a product page with a 360 video as an example of a page that is not one.³¹ An AI clip on a product page or blog post can improve the page for visitors, but it will not earn a video result.
• YouTube ranks the video, whatever made it. YouTube search weighs relevance (the title, tags, description and video content), engagement such as watch time for the query, and quality signals of expertise, authoritativeness and trustworthiness.³² An AI clip earns watch time the same way as any other.
• Volume can count against you. Google Search has no rule against AI content, but generating many pages without adding value can breach its scaled content abuse policy, and AI-generated images on ecommerce sites must carry IPTC DigitalSourceType metadata.³³
• A label does not cost reach. Disclosing AI on YouTube does not by itself limit a video's audience.²²

In my view the bigger weakness of most AI clips for search is how little text comes with them. Search engines and AI assistants work mostly from text: titles, descriptions, captions and transcripts. An eight-second clip with music and no speech gives them almost nothing to read or quote. The chapters on watch pages, YouTube and captions in the image and video SEO guide set out what earns visibility, and measuring visibility in AI search covers checking whether your videos are cited.

In November 2026 it is not a waste of money across the board. For Shorts, b-roll, internal videos and product-led ads it is cheap, and sometimes free. It still wastes money when it is bought to replace brand storytelling, to fake a demonstration, or as a video SEO tactic. Before paying for a plan, I would work through these steps.