Leaked Gemini Omni demos show production-style video edits that rapidly drain user quotas

Video models are starting to look less like toy generators and more like rough-cut editing suites. That is the real story in the Gemini Omni leak. The clips that surfaced this week do not just show “make me a video” prompts — they show targeted edits inside an existing scene, the kind of thing people normally do with layers, masks, and cleanup tools. But the other half of the leak matters just as much: the compute bill looks huge. Early testers say just two generations nearly wiped out a full Google AI Pro daily allowance. ### What actually leaked? Two things leaked. First, TestingCatalog spotted Gemini app strings in early May pointing to a video feature “Powered by Omni,” sitting next to references that looked tied to Google’s existing Veo pipeline. Then, on May 11, demo clips and interface snippets started circulating that showed an unreleased Gemini Omni model generating and editing video inside chat. Google has not formally announced Omni yet. (testingcatalog.com) ### Why are people treating this as a big step? Because the demos look less like raw generation and more like production-style revision. The leaked examples show coherent object swaps, style changes, and cleanup-style edits while keeping motion and scene structure relatively stable. That is the hard part. Anybody can make a flashy four-second clip. Keeping continuity while changing one thing inside a moving shot is where these systems usually fall apart. (testingcatalog.com) ### Is this Veo 4 or something else? Nobody outside Google seems to know yet. The best read from the leaks is that Omni is either a new Gemini-facing video model, a chat wrapper around a newer Veo generation, or a product name for a broader editing workflow rather than a single base model. Metadata and UI placement have pushed a lot of people toward the “next Veo under the hood” theory, but that is still inference, not confirmation. (chromeunboxed.com) ### Why does the quota detail matter so much? Because it tells you where the bottleneck still is. One reported test burned roughly 86% of a daily AI Pro quota after just two video generations. That means the model may be good enough to feel useful, but still expensive enough that Google has to meter it aggressively. In other words, the demos are exciting, but the leak also screams scarcity. (testingcatalog.com) ### Why is editing harder than generating? Think of generation as inventing a short dream from scratch. Editing is more like replacing one actor in a moving shot without changing the lighting, camera path, shadows, or timing. The model has to preserve everything you did not ask to change. That is why watermark removal, object replacement, or scene cleanup demos get attention — they test consistency, not just imagination. (androidauthority.com) ### What does this mean for Google’s AI plans? It suggests Google wants Gemini to be more than a chatbot with media add-ons. If Omni ships near Google I/O 2026, Gemini starts looking like a front door for multimodal creation — text, image, video, and iterative edits in one thread. That would put Google more directly into the same user workflow battle that OpenAI, Runway, and others are chasing. (chromeunboxed.com) ### So what is the catch? The catch is that frontier video still does not look cheap, instant, or abundant. Even if the quality is real, heavy quota burn means most people will hit limits fast, and platforms will keep the best tools behind paid tiers or credits. The leak shows capability, but it also shows why these features are still premium. (testingcatalog.com) ### Bottom line? Gemini Omni looks important because it hints at a shift from AI video generation to AI video editing inside chat. But the leak also reveals the current tradeoff in plain view — better control and coherence, paid for with brutal compute costs and tight caps. (androidauthority.com)