Descriptions:
Google’s next-generation video model, internally referred to as Gemini Omni, was accidentally surfaced to a small group of users before any official announcement, triggering a wave of analysis after one user spotted the label “powered by Omni” in the Gemini app’s video generation tab. The accidental access produced two public video generations before Google apparently corrected the exposure — enough for the AI community to begin benchmarking the model against current competitors.
TheAIGRID places the leaked outputs side-by-side with ByteDance’s Sea Dance 2 using identical prompts, including a professor writing a trigonometric proof on a chalkboard and the informal “Will Smith spaghetti” scene benchmark. The verdict: Omni appears broadly competitive with Sea Dance 2, with output quality that visibly exceeds the current production model (V3.1, code-named Toucan internally). The channel draws a conceptual parallel to OpenAI’s GPT-4o launch, speculating that “Omni” signals a true multimodal architecture capable of accepting and generating across text, audio, image, and video — a meaningful step change rather than an incremental update.
The most concrete data point from the leak is resource consumption: the two generated videos consumed roughly 86% of a user’s monthly allowance on the $20/month Gemini Pro plan — far higher than V3.1 or Sora 2. This implies an architecture significantly larger and more compute-intensive than anything Google has previously offered consumers, with usage limits almost certain to constrain casual creators on standard subscription tiers.
📺 Source: TheAIGRID · Published May 12, 2026
🏷️ Format: News Analysis
