Leaked Gemini 'Omni' demos show realistic short-video generation

Google looks like it’s about to make Gemini much more than a chatbot. Leaked screens and demo clips from May 11 show an unreleased model called Gemini Omni generating and editing short videos inside the Gemini app. The demos look surprisingly polished — more coherent motion, better lighting, fewer obvious AI tells. And because Google I/O starts May 19, this feels less like a random bug and more like a preview that escaped early. ### What actually leaked? Users spotted a new card inside Gemini labeled “Create with Gemini Omni,” plus interface text suggesting templates, remixing, and in-chat video edits. That matters because this was not just a stray codename buried in app code — it appeared in a live product surface, with generated examples attached. Android Authority also surfaced sample outputs and metadata pointing to Omni as a video model tied to Gemini’s existing creation flow. (9to5google.com) ### Is Omni a brand-new model? Maybe not in the clean-room sense. The strongest read from the leak is that Omni builds on Google’s Veo video stack rather than replacing it. Android Authority said metadata points to Veo under the hood, and other leak roundups describe Omni as the Gemini-facing layer for generating, editing, and remixing clips from chat. Basically, Veo may be the engine while Omni is the product experience people actually touch. (9to5google.com) That’s an inference, but it fits the evidence we have so far. ### Why are people focusing on the demo quality? Because short AI video is easy to fake in cherry-picked clips, but these examples seem to clear a higher bar. The leaked outputs showed more realistic movement and scene consistency than the rough, floaty look people still associate with consumer video generators. That does not prove Omni is broadly reliable. But it does suggest Google is getting closer to “type a prompt, get a usable clip” instead of “type a prompt, get a weird experiment.” (androidauthority.com) ### What’s the catch? Compute. One reported generation consumed 86% of a daily Google AI Pro quota. That is the loudest detail in the whole leak, because it tells you this is still expensive enough that Google may need strict caps even for paying users. Good video models are basically image models multiplied by time — every extra second means more frames, more consistency work, and more chances for the whole thing to break. (9to5google.com) ### Why leak into Gemini now? Because Google has been steadily turning Gemini into the front door for everything. The same stretch of reporting shows Gemini for Google Home loosening up enough to answer cocktail-recipe requests, a Gemini overlay rolling out in the Google app, and Google Maps code hinting at Gemini support in CarPlay. Put that together and the pattern is obvious — Google wants Gemini sitting on top of search, home, maps, and now media creation. (androidauthority.com) ### So what changes if Omni launches at I/O? The big shift is interface, not just capability. Google already had impressive video tech in demos, but Omni suggests that video generation could become a normal Gemini action, like asking for a summary or an image. Once that happens, the competitive frame changes. Google is no longer just showing model research — it is productizing video inside an app millions already use. (engadget.com) ### Why does that matter beyond Google fans? Because the AI race is moving from “who has the coolest lab demo” to “who can hide the complexity.” If Gemini can generate, edit, and remix video in one chat flow, then the real product is convenience. People do not want a model zoo. They want one box that understands the request and makes the thing. ### Bottom line? Omni still looks leaked, limited, and expensive. But the direction is clear — Google wants Gemini to be where video generation lives, not just where you ask questions. (9to5google.com)