Seven Secret Gemini Live Models Surface Ahead of I/O
A hidden model selector discovered in the Google app has revealed that Gemini Live is quietly running on a much richer backend than users see today. Buried behind a server-side flag, the menu exposes seven AI model 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 models designed to process speech directly, rather than relying on text intermediaries. Two entries, including the thinking variant, appeared overnight on May 8, indicating they are nearing production readiness rather than being purely experimental. Early testing confirms these models behave differently in practice, with varying access to live weather data and distinct conversational styles. Because the list is delivered from Google’s servers, the company can swap or add models instantly, setting the stage for a flexible, multi-model Gemini Live reveal at I/O.
Inside the Thinking Variant and Emerging AI Reasoning Capabilities
Among the hidden models, A2A_Rev25_RC2_Thinking stands out as Google’s clearest attempt to brand a Gemini Live model around deeper reasoning. While details remain scarce, the dedicated label and RC2 tag suggest a near-final build optimized for complex, multi-step thinking rather than just fluent conversation. Tests across the hidden lineup already show meaningful behavioral differences: some models can pull live weather data, while others cannot, implying different capability tiers and tool access. The Capybara model even identifies itself as “Gemini 3.1 Pro,” signaling a more advanced reasoning core compared with the Flash Live baseline. In parallel, early access to the Gemini Omni video model shows it handling sophisticated prompts such as professors constructing trigonometric proofs, albeit with the familiar visual glitches that still plague AI video. Together, these developments point toward Google prioritizing AI reasoning capabilities as a central differentiator for Gemini Live.
Personalization, Memory, and the Push Toward Agentic Gemini
Beyond raw reasoning, Google appears to be tuning specific Gemini Live variants for long-term usefulness. The P13n (personalization) model behaves differently from the current default, asking which time zone the user is in instead of assuming, and naturally reusing personal details shared earlier in the session. That contrasts sharply with the default Gemini Live model, which resists persistent memory, and hints at a future where Gemini acts more like an adaptive assistant than a stateless chatbot. This aligns with Google’s broader agentic agenda. Gemini Intelligence, announced recently, is designed to automate tasks across apps and the web, with Chrome auto-browse arriving soon. Combined with the Gemini Enterprise Agent Platform, Google is quietly assembling the pieces for agents that can listen continuously, remember user context, and execute multi-step workflows, rather than merely answering prompts in isolation.
Gemini vs ChatGPT and Claude: Closing the Capability Gap
The hidden Gemini Live lineup lands just as Google faces intense pressure from rivals. Reports suggest Google is preparing a new Gemini model comparable to an upcoming GPT-5.5‑class release from OpenAI, though still slightly behind Anthropic’s frontier Mythos model. Yet Google’s problem is less about missing a single leaderboard and more about habit. For many power users, ChatGPT and Claude are already the default mental shortcuts for coding, research, and quick reasoning checks. Gemini has to be more than another capable model: it must feel like the most obvious first stop. A thinking variant for Gemini Live, especially if paired with better tools and reliability, is Google’s attempt to narrow the Gemini vs ChatGPT gap by excelling at sustained reasoning and complex tasks instead of chasing benchmark scores alone.
Winning Developers with Agentic Work and Multi-Model Flexibility
Developers will be the harshest judges of Google’s new Gemini Live strategy. Coding has become the pressure test where AI models either integrate into daily workflows or get relegated to occasional use. Google’s own messaging acknowledges that a model must reduce cleanup, survive messy real projects, and lower friction for agentic work to win adoption. The hidden multi-model selector suggests Gemini Live could dynamically route requests to specialized models—reasoning, personalization, or native audio—while Google’s agent platforms handle orchestration, identity, and security. If Gemini can reliably write and refactor code, manage multi-step tasks, and recover gracefully from vague or bad inputs, it may finally disrupt developer routines built around ChatGPT and Claude. I/O now looks like the moment when Google must prove that a thinking variant and agent-ready Gemini Live can turn impressive demos into indispensable infrastructure.