The Pulse
Google gives live voice agents two different speeds
Google’s Live API documentation distinguishes Gemini 3.8 Live from Gemini 3.8 Live Extended Thinking, outlining different approaches to latency, reasoning, tool calls and session state.
AI.info Team ·
Google gives live voice agents two different speeds
Google’s AI for Developers documentation describes two real-time voice models: gemini-3.8-live and gemini-3.8-live-extended-thinking. The documentation separates direct, low-latency voice exchanges from interactions that require planning, multi-step reasoning or tool calls that may take longer to complete.
Both models are used with the Gemini Live API, which Google describes as supporting real-time, bidirectional voice conversations. The model pages list text, images, audio and video as supported inputs, with text and audio outputs.
“The model plans and calls asynchronous tools in the background while speaking natural conversational fillers to keep the interaction active.” — Google AI for Developers, documentation publisher
The standard gemini-3.8-live model is intended for voice agents where immediate turn-taking is important and tasks are direct. Google lists customer-service triage, language practice, voice search and interactive storytelling among the suggested uses. It also identifies quick external actions, such as reading sensor values or controlling smart devices, as appropriate workloads.
For the standard model, each user turn produces a single model response. A turnComplete: true signal indicates that the model has finished speaking and that the session is idle. Google says the model uses an interleaved reasoning approach with a fixed latency profile; it does not support a configurable thinking_level.
Extended Thinking changes the session lifecycle
gemini-3.8-live-extended-thinking is aimed at voice agents that must inspect complex information, plan several steps or wait for external tools. Its documented uses include support agents assessing logs, error codes and configuration checks; travel and booking agents searching flights and hotels; and tutoring systems that verify formulas or debug code before responding.
Instead of leaving a caller waiting silently while a tool runs, the extended-thinking model can issue spoken interim updates. Google’s documentation uses the example, “Checking flight options now,” while describing this behavior. The model can continue processing in the background, make asynchronous tool calls and provide a final spoken answer when the larger task is complete.
That behavior means turnComplete: true has a narrower meaning in an extended-thinking session. It can mark the end of an individual utterance, including an interim spoken update, without indicating that all reasoning and tool work has ended. Clients must also monitor interaction_status.
An IN_PROGRESS status means the server is still processing the user’s request, conducting background reasoning or awaiting an asynchronous tool response. An IDLE status means that processing, reasoning and tool calls have finished, and the session is ready for further user input. Applications that display a listening state or decide when to accept another request therefore need to use the interaction status rather than rely only on the turn-completion signal.
Different configuration and tool requirements
Google allows developers to configure the extended-thinking model’s reasoning depth with thinking_level values of low, medium or high. The MINIMAL setting is not supported. Function declarations for this model must use non-blocking behavior, NON_BLOCKING; Google says synchronous blocking functions return an error.
That requires clients to handle multiple messages during what a user experiences as one request. A session may receive an interim audio response, a tool call, a tool response and a final answer. The documentation says clients should keep listening after a turnComplete: true signal if interaction_status remains IN_PROGRESS.
Google’s migration guidance identifies gemini-3.1-flash-live-preview as an earlier Live API model. Moving an existing application to gemini-3.8-live requires changing the model string and omitting thinking_level or thinking_config from the standard model’s setup. Applications adopting extended thinking must add status handling, configure non-blocking tools and account for intermediate spoken output.
The distinction leaves developers with a routing choice: use Gemini 3.8 Live for fast conversational turns and direct commands, or use Gemini 3.8 Live Extended Thinking when a voice interaction needs deeper reasoning and longer-running external work.