Google ships Gemini 3.8 Live with background tool calls during voice conversations

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September for real-time voice applications. The main model can accept text, image, audio and video input, maintain spoken dialogue and issue asynchronous function calls so a client can execute a tool or API request while the model continues streaming audio to the user.

Google also says Gemini 3.8 Live can use live visual context and move between 97 supported languages during a conversation. Those capabilities are available to developers through the Gemini Live API and Google AI Studio rather than being limited to a single Google consumer application.

Gemini 3.8 Live Extended Thinking is aimed at more complex, multi-step tasks. Google describes it as able to reason while continuing the spoken interaction, including giving progress narration while background work proceeds. The practical change is that a voice agent no longer has to become silent every time a tool call or longer reasoning step begins.

The standard 3.8 Live model is also rolling into Search Live. Extended Thinking is rolling out in Gemini Live and selected Workspace experiences, while enterprise access begins as a private preview. Availability therefore varies by product even though both models are already exposed to developers through the API.

Google lists audio input at $0.005 per minute and audio output at $0.018 per minute for the new Live models. Its developer documentation adds an important constraint: Live API sessions are billed by tokens in the active context window, and earlier context can be processed again on later turns. Long conversations can therefore cost progressively more unless applications use context-window compression or otherwise limit retained history.

Google’s Live API documentation also warns that native audio remains in conversational history as audio tokens and that transcription adds separate text-output charges when enabled. Developers building long-running voice experiences therefore need to account for context growth, transcription and listening time rather than extrapolating total cost from the headline per-minute audio rates.

Google reports strong results on speech-agent and task-completion benchmarks, but those figures come from Google and the benchmark providers under specific evaluation setups. They do not establish how reliably asynchronous tools, interruption handling or long-running voice sessions will behave across production applications.

The more consequential evidence will be operational: tool-call success rates, latency under load, conversation recovery after failed tools and the real cost of long sessions. The release confirms that background execution during live dialogue is now a first-party Gemini capability; its reliability and economics still need broader independent testing.