Google Tests Multiple Hidden AI Models for Gemini Live

Google is internally testing seven new AI models for its Gemini Live voice feature, uncovered through a hidden menu in the latest version of the Google app. The models produce measurably different responses in conversation tests, with four able to access the user’s location for live weather updates while three cannot.

One model identified itself as Gemini 3.1 Pro rather than the current Gemini 3.1 Flash Live version used for voice chats. Testing also showed differences in memory and personalization. Three models promised to retain personal details shared earlier in conversations, while others refused.

Every new model outperformed the current default in referencing prior chat details. The hidden selector appeared in Google App version 17.18.22 behind a server-side flag. It lists models including variants labeled with terms suggesting audio-to-audio processing, personalization, experimental versions and a specialized “Thinking” model with enhanced reasoning.

Two models reached Release Candidate 2 stage as of May 8, indicating they are nearing production readiness. The company can add or remove models via its servers without requiring an app update.

When asked to identify itself, one model responded with a more powerful designation than the standard Flash Live version currently deployed. Some models accepted a deliberately false claim during testing while two others, including a nitrogen-labeled variant, detected the inaccuracy.

The infrastructure suggests the company is evaluating specialized voice AI experiences. Users could eventually choose between faster responses and more thoughtful ones that take additional time, or access stronger capabilities through a paid tier. Google I/O 2026 is scheduled to begin May 19.

The presence of near-production models and a functional selector interface points to potential announcements around expanded voice AI options at the event. Criminal groups and state-linked actors from China, North Korea and Russia are using commercial models including Gemini, Claude and OpenAI tools to refine and scale attacks.

The actors employ the models to boost speed, scale and sophistication, enabling better malware development, vulnerability testing and persistence against targets. A criminal group was recently on the verge of leveraging a zero-day vulnerability for a mass exploitation campaign using an AI large language model.

Anthropic declined last month to release its Mythos model after determining it could discover zero-day vulnerabilities in every major operating system and web browser, posing risks to governments and financial institutions if misused.

““There’s a misconception that the AI vulnerability race is imminent. The reality is it’s already begun.””

Security experts note AI tools can assist both attackers and defenders. One professor of security engineering said the field has shifted to LLM-assisted bug discovery, with consequences still unfolding. A separate report questioned assumptions about AI-driven productivity gains in the public sector, citing gaps between projected benefits and measured outcomes.

The company continues testing the voice models as Google I/O approaches, with the selector currently limited to internal use. Further updates to the model list or interface could appear before or during the conference.