AI video moved faster in 2026 than any other part of generative AI. Fifteen- to thirty-second clips with synchronised dialogue and sound effects are now normal, editing tools let you change a finished shot with a sentence, and some of the strongest models come from China. Meanwhile the product many people thought would lead — OpenAI's Sora — was shut down. Here is where things stand in October 2026.
First: what happened to Sora?
OpenAI announced in March 2026 that it was "saying goodbye to the Sora app", and the app and website were discontinued in April, less than seven months after launch (TechCrunch). Disney's planned licensing deal with OpenAI for Sora fell through with it. OpenAI has not announced a replacement video product. If you relied on Sora, the tools below are the realistic alternatives — and the episode is a good argument for building workflows that are not tied to a single vendor (Forbes).
The leading AI video generators in 2026
Google Veo 3.1 and Gemini Omni Flash
Veo 3.1 remains Google's dedicated video model — there is no Veo 4 yet. A January update added native vertical 9:16 video for Shorts and Reels and upscaling to 1080p and 4K (Google). At I/O in May, Google also introduced Gemini Omni Flash, a model that takes any mix of text, images, audio and video and produces video, with conversational editing over multiple turns; clips were capped at 10 seconds at launch and every output carries a SynthID watermark (TNW). It is available in the Gemini app, Google's Flow filmmaking tool and the Gemini API.
Best for: teams in the Google ecosystem, YouTube Shorts creators and developers who want video through the Gemini API.
Kling 3.0 (Kuaishou)
Kling 3.0 launched globally in February with up to 15-second clips, native audio across multiple languages, dialects and accents, and multimodal input (Kuaishou via Bernama). It is also a commercial success: Kuaishou reported Kling revenue of more than RMB 850 million in Q2 2026 (Kuaishou). Kling is available inside Adobe Firefly and Runway as well as its own apps.
Best for: realistic people and motion, multilingual dialogue, and high-volume social content.
Runway Gen-4.5 and Aleph
Runway's flagship is still Gen-4.5 — there is no Gen-5 — but the company shipped a stream of tools around it in 2026: Aleph 2.0, which propagates an edit across a whole video; real-time conversational avatars; an agent for multi-step video tasks; and plugins for Premiere Pro, After Effects and DaVinci Resolve (Runway changelog). Runway now also hosts third-party models such as Kling, Seedance and Gemini Omni, making it a one-stop studio.
Best for: professional editors and agencies who want AI inside an existing post-production workflow.
ByteDance Seedance 2.0 and 2.5
Seedance 2.0 launched in February with joint audio-video generation and up to 12 reference files for consistent characters and scenes (eWeek). After Disney, Netflix and other studios sent cease-and-desist letters, ByteDance paused the global rollout in March (TechNode); it later reached CapCut in some markets and Runway outside the US. Seedance 2.5, launched in the summer, generates clips of up to 30 seconds in a single pass with native audio (TechTimes).
Best for: longer single-shot clips and reference-driven consistency — check availability and licensing terms in your market.
Luma Ray3.2
Luma's Ray3.2, released in June, is its first model offered through an API, with clips up to 20 seconds at 1080p, up to 16 keyframes for precise control, facial-performance tracking and native HDR output with 16-bit EXR export (Luma).
Best for: film and VFX-style work where colour and frame control matter.
Alibaba Wan
Alibaba's Wan 3.0, in public beta since August, produces up to 30-second 1080p clips with multilingual voice — but unlike earlier versions its weights are not open (TNW). Wan 2.2, released under the Apache 2.0 licence in 2025, remains the strongest open-weight option for teams that need to self-host video generation.
Best for: Wan 2.2 for self-hosting and private pipelines; Wan 3.0 via Alibaba Cloud for longer clips.
Also worth knowing
- Adobe Firefly — Adobe's AI assistant works across Creative Cloud and gives access to more than 30 models, including Veo, Kling, Runway and Luma, with Adobe's commercial-use terms (Adobe).
- Midjourney — now on its V8 image models, with image-to-video for short animated clips.
- xAI Grok Imagine 1.5 — added image-to-video with native audio in June (The Decoder).
- Vidu (Shengshu) — released its Q4 preview in October, priced from $0.014 per second (PR Newswire).
- Avatar tools — Synthesia and HeyGen turn scripts into presenter videos for training, sales and localisation.
Quick comparison
| Tool | Max clip (approx.) | Native audio | Stand-out strength |
|---|---|---|---|
| Google Veo 3.1 / Gemini Omni | Short clips; Omni 10s at launch | Yes | Google ecosystem, vertical video, conversational editing |
| Kling 3.0 | 15s | Yes, multilingual | Realistic people, dialogue |
| Runway Gen-4.5 + Aleph | Varies | Via models and tools | Editing and pro workflows |
| Seedance 2.5 | 30s | Yes | Long single-pass clips |
| Luma Ray3.2 | 20s | — | Keyframes, HDR, API |
| Wan 2.2 (open) / 3.0 | 30s (3.0) | Yes (3.0) | Open weights (2.2) |
Specifications change frequently; check each provider's current documentation.
How businesses are using AI video
- Ad variations at scale. The most common use among ad buyers is versioning one video for different audiences, formats and placements. US digital video ad spend is forecast to pass $80 billion in 2026 (IAB).
- Product videos for e-commerce from a handful of product photos.
- Localisation — dubbing and lip-synced versions in many languages.
- Training and onboarding with avatar presenters.
- Real estate — listing videos and virtual staging.
- Social content — Shorts, Reels and TikToks produced in hours instead of days.
The market is growing quickly from a small base — Fortune Business Insights estimates the AI video generator market at about $847 million in 2026, rising to more than $3 billion by 2034 (Fortune Business Insights).
What to check before you use AI video commercially
- Licensing and commercial terms — who owns outputs, and does the provider indemnify you?
- Training-data disputes — some models face copyright challenges; prefer providers with clear licensing for high-profile campaigns.
- Likeness and consent — never generate identifiable real people without written permission.
- Labelling — EU AI Act transparency rules apply from August 2026, and YouTube and TikTok label AI content. See what is AIGC for the rules.
- Vendor risk — Sora's shutdown shows why you should keep prompts, assets and workflows portable across models.
- Audience reaction — obviously synthetic ads can backfire; use AI where it improves the work, with human creative direction.
Building AI video into your product
If you run a platform — e-commerce, real estate, education, marketing software — the bigger opportunity may be letting your users generate video inside your product: product clips from listings, narrated lessons, localised marketing. Zinetics builds these features on top of video, image and voice APIs, with moderation, labelling, provenance and cost controls. See our AI development services or talk to us.