The market after the Sora shutdown
Until spring 2026, Sora was the face of the AI video wave. Then OpenAI pulled the plug - the app and web version have been offline since late April, the API follows on September 24, 2026, and no successor has been announced. The full story including detection cues is in our guide to Sora 2 videos.
Demand did not shrink, it spread out. Three names define the year: Veo, Kling and Runway. Add cheap volume providers like Hailuo and freely available models like Wan that anyone can run on their own hardware. Realism has gone up everywhere - the days when six fingers were a reliable tell are coming to an end.
Google Veo: the front-runner with a mandatory watermark
Veo is currently the benchmark for believable footage with matching sound: version 3.1 produces clips with a native audio track, sticks closely to the prompt and delivers 4K on demand. Which is exactly why Veo clips flood Shorts and Reels.
For detection, Veo is the most pleasant case: Google embeds an invisible SynthID watermark into every generated video - directly into the pixels of every frame. The catch: in practice only Google's own tools can read that signal, and depending on the product a visible logo is added on top, which can be removed like any visible logo. How invisible watermarks work in general is covered in our guide to C2PA, SynthID and watermarks.
Kling: the price breaker from China
Kling, by Chinese provider Kuaishou, worked its way up through pricing: a lot of video length for a few cents per second. Version 3.0, released in early February 2026, manages clips of up to 15 seconds in native 4K at 60 frames per second - technically that puts Kling in the top group, and in our checking practice Kling clips show up accordingly often.
On labeling, Kling is far less dependable than Veo: whether a visible logo or machine-readable provenance data comes along depends on the product and the export path. In practice that means a Kling video in your feed often carries no hint of its origin at all.
Runway: the professionals' tool
Runway targets people who produce video rather than scroll it: advertising, music videos, film previsualisation. The Gen-4 model family - most recently in its Gen-4.5 version - regularly sits at the top of independent comparisons, and the real edge is the workflow - precise control over camera, characters and scenes instead of pure prompt roulette.
Good news from a detection standpoint: Runway is one of the providers that attach C2PA provenance data to their output (Adobe does the same for its Firefly content) - exactly the cryptographically signed credentials our check reads. But the old rule applies here too: re-uploads and editing often strip such metadata.
The open models: strong, free, unmarked
Alongside the big providers there is a growing family of freely available video models, led by Wan from Alibaba. Anyone can download them and run them on their own hardware - no content rules, and no watermark or provenance data out of the box; whether an output gets marked depends entirely on whoever operates the model. Quality-wise they trail Veo and Kling, but the gap shrinks with every release.
For scammers and disinformation these models are the first choice, precisely because nothing is enforced technically and no provider can block abuse. Legally, the AI Act's labeling obligation does apply to them too - in practice, whoever wants to deceive simply does not label.
What does this mean for detection?
Remember the map like this - it describes the original outputs, and sharing or editing strips much of it: Veo videos carry an invisible SynthID signal (which outsiders can barely verify), Runway and Adobe output carries C2PA provenance data (which we read), the legacy Sora stock carries C2PA plus the moving logo - and open models carry nothing. A missing label says little; a present one says a lot.
The checking order therefore stays the same for every generator. First: clarify the source - who posted the video first? Second: look closely - the typical signs of AI videos such as inconsistent cuts, dissolving text and almost-synced lips apply across generators. Third: get a technical check - we read provenance data, look for forensic traces and say honestly when the evidence does not support a verdict.