Seven Secret Gemini Live Models Emerge Inside the Google App
A hidden model selector tucked inside Google App version 17.18.22 has exposed a quietly expanding roster of Gemini Live models. Discovered by Forbes contributor Paul Monckton, the menu—protected by a server-side flag—lists seven distinct AI options: Default, A2A_Rev25_RC2, A2A_Rev25_RC2_Thinking, A2A_Rev23_P13n, A2A_Nitrogen_Rev23, A2A_Capybara, A2A_Capybara_Exp, and A2A_Native_Input. The A2A label strongly suggests “Audio-to-Audio,” indicating models that process speech directly instead of routing everything through text. Two entries, tagged RC2, appeared overnight on May 8, signaling that at least some variants are nearing production readiness rather than being mere prototypes. Monckton’s tests showed each model producing measurably different responses and levels of access to live data such as local weather, confirming these are not just cosmetic labels but meaningfully distinct Gemini Live models under active development as Google refines its AI portfolio.
A Thinking Variant and Personalization Hint at a More Capable Gemini
Among the hidden Gemini Live models, the standout is a dedicated thinking variant, A2A_Rev25_RC2_Thinking, apparently tuned for more complex reasoning tasks than the default live model. Monckton’s twelve-test run across variants showed consistent behavioral differences, with some models capable of pulling real-time location data and others not. The P13n (personalization) model, A2A_Rev23_P13n, was particularly revealing: instead of assuming a time zone, it asked the user to specify one, then remembered personal details for later use in conversation—behaviour the current Gemini Live default refused to replicate. Another model, Capybara, identified itself as “Gemini 3.1 Pro,” differentiating it from the Flash Live model Google typically deploys. Together, these signs suggest Google is experimenting with tiered AI model capabilities, from faster baseline responses to deeper, more context-aware reasoning that can better rival ChatGPT and other advanced AI assistants.
Multi-Model Strategy: One Platform, Many Gemini Live Personalities
The emerging picture is a multi-model Gemini Live platform designed to match different user needs and developer scenarios. Because the model list is delivered from Google’s servers rather than baked into the app, Google can swap or add models on the fly—ideal for live experiments and staged feature rollouts. Four of the hidden models accessed location-based weather data while three did not, implying Google can tune privacy, latency, and feature depth per model. A2A_Native_Input hints at direct handling of raw audio, while experimental labels like Capybara_Exp suggest rapid iteration. For developers, this architecture promises more precise targeting of AI model capabilities, from lightweight real-time voice support to heavier-duty, thinking variant AI for complex reasoning. For users, it sets the stage for a more personalized and adaptive AI experience in which Gemini Live can effectively shape-shift based on context, task complexity, and resource constraints.
Gemini Omni and the Video Frontier Raise the Stakes
Parallel to the Gemini Live experiments, Google is testing Gemini Omni, a next-generation video-focused AI model surfaced via an early-access pop-up inviting users to “create with Gemini Omni.” Metadata ties Omni to Google’s Veo foundation, and early tests show it handling sophisticated prompts like a professor writing a trigonometric proof on a chalkboard with lifelike visual results. Yet the model still shows familiar AI video glitches, such as spontaneously appearing spaghetti in a dinner scene, and strict guardrails reportedly blocked running the meme-famous “Will Smith eating spaghetti” benchmark. Critically, Omni appears extremely resource-intensive: a user hit 86% of their daily usage limit after just two video generation requests on the Google AI Pro plan. This heavy compute footprint helps explain why Google is adding more explicit usage limits and underscores how video-native AI will be a key battleground in the broader ChatGPT competitor landscape.
I/O Timing and Competitive Pressure in the AI Model Race
The timing of these discoveries is no coincidence. With Google I/O scheduled to begin on May 19, the presence of RC2-tagged Gemini Live models and a server-driven selector strongly suggests that switchable voice and reasoning modes could feature in live demos. Combined with Gemini Omni and the newly announced Gemini Intelligence—Google’s system for automating tasks across apps and the web, including Chrome auto-browse arriving in June—Google appears to be laying the groundwork for an integrated, multi-modal AI ecosystem. Strategically, the hidden Gemini Live models signal Google’s intent to close capability gaps with ChatGPT and other AI platforms by offering a spectrum of AI model capabilities rather than a single monolithic system. If executed well, this multi-model approach could turn Gemini Live into a flexible AI front end that adapts to use case, cost, and complexity, reshaping how developers and users evaluate ChatGPT competitors.