Google releases Gemini 3.8 Live for conversational voice agents
The new Live models support background tool calls while a conversation continues, with a separate option for deeper reasoning.
THE STORY IN THREE POINTS
- The September 15 announcement introduces Gemini 3.8 Live models.
- Live Extended Thinking adds configurable reasoning.
- The models support asynchronous function calls during audio streaming.
The model release
Google announced Gemini 3.8 Live and Live Extended Thinking for developers through the Gemini API and AI Studio. The models are designed to perform tasks while maintaining a spoken conversation.
Capabilities described include asynchronous tool calls, live visual context and configurable thinking in the extended variant. The same post discusses the previously released Gemini 3.5 Transcribe model; the two products serve different voice and transcription workflows.
Editorial context: conversation has a timing problem
A voice agent may need several seconds to finish a tool call while the person keeps talking. A useful interaction should make progress understandable and avoid treating a provisional result as a confirmed action.
An evaluation could include interruptions, corrections and a change to the requested destination before execution. Those moments would reveal whether the system maintains the user's latest intent when conversation and background work proceed together.
Editorial outlook: what to watch
Developers should compare complete interactions, including recovery from failed calls and misunderstood names or numbers.
The quality of a spoken answer is only one part of a task-oriented voice product. The action it triggers, its confirmation and the user's ability to correct it deserve equal attention.
Sources: Google: Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe ↗
