Google surfaces seven Gemini models

Google’s Gemini story this week is really two leaks that fit together. One is about voice — seven hidden Gemini Live models sitting behind a server-side flag in the Google app. The other is about video — an unannounced “Gemini Omni” mode that can apparently create and edit clips inside chat. Put them together and the shape of Google’s next move gets a lot clearer: Gemini is turning into a model menu, not a single assistant. ### Where did the seven models show up? They showed up in the Android Google app, version 17.18.22, where a hidden settings button exposed a model selector for Gemini Live. The leaked options included codenames like “Capybara” and “Nitrogen,” plus thinking and personalization variants. That matters because Gemini Live has mostly felt like one voice mode with one personality. The hidden selector suggests Google is already testing multiple back-end behaviors for the same voice surface. (forbes.com) ### Were these actually different models? Looks like yes — not just different labels. Early testing found meaningful behavior differences between the options. Four models could access location and return live weather, while three could not. One model, Capybara, identified itself as “Gemini 3.1 Pro” instead of the usual Flash Live identity. Some variants also appeared better at memory and personalization. (forbes.com) Basically, Google seems to be tuning voice for different tradeoffs — speed, reasoning, memory, and tool access. ### What is Gemini Omni supposed to be? Omni looks like a new video generation layer inside Gemini. A user saw a “Create with Gemini Omni” prompt that described a tool able to remix videos, edit directly in chat, use templates, and do more creative work without leaving the conversation. Metadata pointed to a connection with Veo, Google’s existing video model family, but the interface suggests something broader than a plain text-to-video box. (forbes.com) It looks more like multimodal editing woven into Gemini itself. ### Why is the quota detail such a big deal? Because it turns the leak from demo hype into product reality. Two Omni generations reportedly consumed 86% of a daily AI Pro allowance, even with some same-day Gemini Flash use mixed in. That is a huge hint about compute cost. Google’s own subscriber help pages already show tight daily caps for video generation — 3 videos per day on AI Pro with Veo 3.1 Lite, 5 on AI Ultra with Veo 3.1 Pro — and warn that limits can change. (9to5google.com) Video is expensive, and Google is clearly budgeting it hard. ### Why not just run one model for everything? Because the economics break fast. Voice chat wants low latency. Reasoning wants more compute. Personalization needs memory and permissions. Video generation is in another league entirely. Google’s API docs make the same point in a more mechanical way — limits vary by model, tier, and request type, with stricter caps for preview and experimental systems. In plain English: one “best” model is too slow, too costly, or too constrained for every surface at once. (9to5google.com) ### So what is Google probably preparing for I/O? Most likely a more explicit tiering of Gemini experiences. Not just Basic, Pro, and Ultra pricing, but model choices underneath those plans — fast voice, thinking voice, personalized voice, and premium video generation with stricter quotas. The hidden selector shows the plumbing is already there. Omni shows Google also wants video creation to live inside the same product, even if usage has to be rationed. (ai.google.dev) That is the catch — better multimodal workflows, but only if Google can keep latency and cost under control. ### Bottom line? The leak is less about seven weird codenames than about Gemini’s future shape. Google appears to be moving from one chatbot to a stack of specialized models, each with its own limits, strengths, and price logic. If that holds, the real product decision won’t be “use Gemini or not.” It’ll be “which Gemini do you get, and how often can you afford to call it?” (forbes.com)