Google's Gemini Omni Wants to Be a World Model

At Google I/O on May 19, 2026, the company introduced Gemini Omni, its most ambitious multimodal release so far. The first public version, Gemini Omni Flash, is already rolling out to subscribers through the Gemini app, Google Flow, and YouTube Shorts.

Omni takes text, images, audio, and short video as input and produces video as output. A user can sketch a scene, hum a melody, or type a description, and Omni will try to render the result as a clip of up to ten seconds. Every output carries a SynthID watermark, Google's invisible provenance marker, so downstream platforms can identify the content as synthetic.

The detail Demis Hassabis kept returning to in the keynote was framing. Omni, he argued, is not a video generator. It is a world model, a system that builds an internal representation of physical reality and reasons about what should happen next inside a scene. The distinction matters because it points at a different research trajectory: video as the output of a model that simulates the world, rather than the output of a model that imitates other videos.

That ambition runs into the limits of the first release pretty quickly. Omni Flash caps its clips at ten seconds and is the cheaper, faster tier of the family. The full Omni model, with longer duration and stronger fidelity, sits behind paid Ultra access and developer pipelines that Google has not opened broadly yet. Reviewers who have tested the Flash version describe the output as competitive with Runway and OpenAI Sora on short clips, though weaker on long-horizon physical consistency, which is exactly where a world-model claim would need to deliver.

Google released Omni alongside Gemini 3.5 Flash, the latest workhorse tier of its main reasoning family, and Gemini Spark, the always-on personal agent that runs on Google Cloud virtual machines. Whether Omni earns the world-model label or remains a very polished video tool is a question the next few months of usage will settle.