Video models in 2026: from demo to ad

The verified landscape as of September 2026

Models whose specs and prices we confirmed in each vendor's documentation.

Google now recommends Omni Flash as the default model and keeps Veo 3.1 for scene extension or last-frame control. In the preference rankings (Artificial Analysis on September 26, Arena on the 21st), the top spots belong to Gemini Omni Flash, Wan 3.0, MiniMax H3 and Seedance 2.x; the 2025 leaders (Kling, Runway Gen-4.5, Veo 3.1, Luma and Sora 2 Pro) now sit mid-table.

The Sora shutdown: a lesson in vendor risk

OpenAI launched Sora 2 on September 30, 2025 and opened it in its API a week later. On March 12, 2026 it expanded that API (20 s clips, 1080p, reusable characters) and twelve days later announced it would retire it. The app and web experience closed on April 26 and the API shut down on September 24, 2026, with no replacement. According to TechCrunch, the app was burning about $1 million a day in compute while its users fell below 500,000, and Disney's $1 billion deal collapsed before any money changed hands.

• Model-agnostic workflow: script, storyboard, key frames and prompts are yours; the video engine should be swappable within days.
• A second engine, tested: validate at least two models on your key frames before committing to a campaign schedule.

What works in production and what doesn't

Documented cases show the real scale. For the Kalshi ad aired during the 2025 NBA Finals, made mainly with Veo 3, its creator reported 300 to 400 generations for 15 usable clips, one person, two to three days and about $2,000. Coca-Cola used five specialists, more than 70,000 clips and about 30 days for its 2025 holiday ads, and Svedka spent about four months rebuilding a character for its Super Bowl LX spot (February 8, 2026).

Works today

• Short spots and social content cut from shots a few seconds long.
• Previs and animatics: AMC Networks agreed in 2025 to use Runway to pre-visualize shows.
• Variants at scale: according to Google, WPP uses Veo and Imagen to turn one creative idea into hundreds or thousands of versions.
• Loops for screens and backdrops for live experiences.

Not yet

• Long continuity: each generation lasts seconds, Veo extensions drop to 720p and a stable character takes weeks.
• Exact brand assets: logos, packaging and legal lines are composited in post.
• Real faces without clean clearance: vendors block or restrict them, especially in Europe.
• French voice without testing: Kling 3.0 speaks English, Chinese, Japanese, Korean and Spanish, and Seedance 2.5 advertises 11 languages, Spanish among them.

The main risk is reception. Coca-Cola drew criticism again in 2025, and McDonald's Netherlands pulled an AI-made holiday ad within four days (December 6 to 10, 2025). IAB measures the gap: 82% of ad executives believe young consumers like AI ads, against 45% of those consumers.

A production workflow in seven steps

The one we recommend at Slash to go from demo to an asset you can air.

1. Script: length, formats (16:9, 9:16), with or without sound, markets, languages and what must be real.
2. Storyboard: shots a few seconds long, because no verified model goes beyond 30 s per generation.
3. Approved key frames: generated with an image model and product references; the client approves stills, which cost cents, before paying for video.
4. Image-to-video: each shot starts from its approved frame, with first and last frame when the model allows it, and several takes per shot.
5. Edit: cut, grade and upscale in your editor; logo, packshot and legal lines are composited on top.
6. Sound: native audio is fine for drafts; the final mix uses licensed music and contracted voices.
7. Compliance: consents, licenses, preserved provenance, market-appropriate labeling and archived masters.

Steps 1 to 3 follow the method in image models in 2026, and we cover sound in voice and audio AI.

What it costs: an illustrative calculation

Price per second matters less than the usable-take rate.

Illustrative assumptions: a 30-second spot with 10 shots; each take is generated at 8 s in 1080p and 1 in 20 is kept, the ratio the Kalshi ad's creator reported. Total: 200 takes and 1,600 generated seconds.

The generation bill ranges from about $130 to $640. The multiplier that matters is the usable-take rate: if it halves, the cost doubles. And human work dominates: the Kalshi ad cost about $2,000 in total according to its creator, and McDonald's Netherlands' holiday production kept about 10 people busy for five weeks, according to NBC. For screens without sound, silent Veo 3.1 on Google Cloud costs half.

Faces, voices and consent

Vendors already limit real people, and more so in Europe. In the EU, the UK, Switzerland and the Middle East and North Africa, Veo only generates adults. Gemini Omni Flash does not allow editing uploaded videos in the European Economic Area, Switzerland and the UK, restricts images with minors there and prevents certain recognizable people from appearing in uploaded images.

The law points the same way. In Colombia, Law 2502 of 2025 aggravates impersonation done with AI, and Law 1581 requires authorization to process personal data, a person's image included; in France, influencer advertising with generated faces must carry the mention Images virtuelles. For actors and ambassadors, the contract should name the AI use, media, territories, duration and fee; for the public at an activation, explicit consent and a deletion policy. More on impersonation risk in deepfake fraud.

Article 50: how to label a video in the EU

Since August 2, 2026, whoever publishes a deepfake in the EU must disclose it clearly and at the latest at first exposure (Article 50, paragraphs 4 and 5). The AI Act calls a deepfake content that resembles existing persons, objects, places, entities or events and would falsely appear authentic (Article 3, point 60): a photorealistic video of a real product in a real city can qualify. Providers, for their part, must mark outputs in a machine-readable way; Veo and Omni put SynthID on every clip.

The Code of Practice of June 10, 2026 makes it concrete: an icon or label perceivable at first exposure without user action, for example top right, and for video, at the start and at regular intervals, both online and offline. Evidently artistic or fictional works may carry it in notes or credits, provided it is perceivable at first exposure. On September 24 the Commission published optional icons in SVG and PNG.

It also applies to agencies outside the EU when the video is distributed in the Union (Article 2), with fines of up to €15 million or 3% of worldwide turnover. Details in rights and provenance of AI content.

DOOH: what changes on street screens

Production practices we recommend, not statistics.

• Formats: Veo 3.1 and Omni Flash only output 16:9 or 9:16, and Seedance 2.0 spans 21:9 to 9:16. For panoramic screens or LED walls with unusual sizes, generate in the closest standard format and adapt in post.
• Loops: a street asset repeats, so the start and end must join. Veo 3.1's first-and-last-frame control helps, and Midjourney documents loops in its video model.
• No sound: most street screens play no audio; don't pay for it. Silent Veo 3.1 on Google Cloud and silent Kling 3.0 cost less.
• Legibility: slow motion, high contrast and large text composited in post, for an audience that looks for a few seconds from a distance.
• Operator specs: resolution, frame rate, codec, slot length and content policy vary by network; get them before generating.
• Label: the code requires the signal offline too; in a short loop, the simplest option is an icon that stays on screen throughout.

This is the approach we recommend for DOOH. For audience activations, see generative AI brand experiences.