Google ships Gemini 3.8 Live for real-time voice with mid-call language switching

Google has released two new real-time voice models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, rolling out across the Gemini API, Google AI Studio, Gemini Enterprise, Search Live, Gemini Live and parts of Workspace. The pair splits the job: the base model targets scale and cost efficiency, while Extended Thinking is built for tasks that need more reasoning without breaking the flow of conversation.

What is new for voice interaction

Gemini 3.8 Live processes visual input in near real time, so a conversation can reference what a camera or screen is currently showing rather than a static snapshot. It also detects and switches between 97 supported languages mid-conversation, and can run tool calls and API requests in the background while continuing to talk, acknowledging a request and narrating progress rather than going silent while a task completes. Google says generated audio carries a SynthID watermark.

Numbers to read as vendor claims, not audits

Google cites third-party evaluation figures for Extended Thinking: a score of 82.6 on Artificial Analysis' Speech-to-Speech Quality Index, and 68.6% on the τ-Voice agentic-task benchmark (35.1% on a banking-specific variant). These come from named benchmark providers cited in Google's own announcement rather than an independent audit verified directly here, and they use different measurement methods from each other — they should not be read as one unified performance score.

Where this lands for MSPs

For managed service providers already juggling more than one voice model across customer deployments, the practical question these releases raise is not the benchmark scores but where inference actually runs and which tenant's data touches which region — a routing and isolation question, separate from any model-capability marketing.

Sources: Google Blog, 9to5Google.