Google's Gemini 'Omni' Video Model Leaks Days Before I/O 2026, Pointing to a Unified Multimodal System

Google's Gemini app has surfaced a new video generation model branded "Omni" in test UI strings discovered this week, days before the company's I/O 2026 developer conference on May 19-20. Screenshots and early demos shared on Reddit and X show a "Create with Gemini Omni" entry point inside Gemini's video tab, with the tagline "Start with an idea or try a template. Powered by Omni."

The leaked clips suggest a capability set that extends well past current text-to-video tools. Users have generated scenes — a professor working through a chalkboard proof, two diners eating spaghetti at a seaside restaurant — but the more interesting demos showed Omni editing existing footage directly through chat: swapping objects in a clip, removing watermarks, and rewriting scenes by typing instructions. Both generation and editing in a single conversational interface would be a first for any top-tier video model.

Metadata buried in the Gemini app hints that Omni is "an extension of Veo," Google DeepMind's existing video generator, rather than a clean-sheet model. That points to one of three interpretations the AI press has floated: not a rename of the Veo pipeline, but a new DeepMind-trained model that sits alongside Veo while sharing some of its tooling. A third, more ambitious reading — that Omni is a unified multimodal system handling text, image, and video in one stack — is plausible but unconfirmed.

The timing is unsubtle. Google has explicitly teed up Gemini updates at I/O 2026, and the company has been on the back foot in generative video since ByteDance's Seedance 2 launched with state-of-the-art cinematic quality. Early Omni samples reportedly lag Seedance on raw generation fidelity but stand out for in-context editing, the area Google can credibly differentiate on given its Gemini chat surface. OpenAI shut down its competing Sora platform in March, leaving the high end of consumer video AI to Google, ByteDance, Runway, and Luma.

One curiosity from the leaked tests: two prompts reportedly burned 86% of a Gemini AI Pro user's daily Omni quota, suggesting either an extremely heavy model or aggressive pricing tiers at launch. Either way, the May 19 keynote is where Google will fix the story — and where the company finds out whether it can turn a leak-driven preview into a clean launch narrative against rivals who have been shipping for months.