Google launched two Gemini 3.8 Live voice models that reason across 97 languages

Google published release notes for Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two voice models designed to talk, parse visual inputs, and execute software actions simultaneously.

The standard Live version handles real-time dialogue and visual context at lower operating costs, offering automatic switching across 97 languages mid-speech. The Extended Thinking model targets complex agentic workflows, using spoken status updates to explain its steps while making background calls to external services without stopping the conversation.

In company benchmark reports, Gemini 3.8 Live Extended Thinking recorded an 82.6 rating on the Artificial Analysis Speech to Speech Quality Index and reached 97.7% on Big Bench Audio. On tests designed for multi-step agent actions, it completed 68.6% of tasks on the general τ-Voice benchmark and 35.1% on a Sierra benchmark tailored to banking workflows.

Access remains split by customer tier. Enterprise agent features are currently restricted to private previews, several Workspace integrations require paid subscription plans, and the system failed roughly two-thirds of the trials on the banking agent test.