A voice model designed to keep talking
Google has introduced two Gemini 3.8 Live models aimed at conversations that continue while the assistant reasons or uses tools. The Google announcement, dated 15 September, describes Gemini 3.8 Live as the lower-latency option and Live Extended Thinking as the higher-reasoning option for more complicated tasks. Google’s developer release makes both available through the Live API, while appearances in consumer and workplace products follow their own rollout paths.
The distinction matters because a spoken assistant is not simply a text chatbot with speech added at the edges. It must respond to interruptions, maintain a coherent exchange and decide when to ask for more information. A slow tool call should not leave a user wondering whether the system heard the request. Google says these models can continue streaming speech while tool calls run in the background.
What changes in the interaction
Gemini 3.8 Live can use near real-time visual input alongside a voice conversation. That could let a user point a camera at an unfamiliar control panel or form and ask a question without first describing every visible detail. The model also supports switching among 97 languages during a conversation, according to Google. That is a capability claim, not a guarantee of equal quality for every accent, task or language pair.
For developers, asynchronous function calling is an important addition. A voice agent might acknowledge a request, look up a booking or check a database, and keep the conversation alive while it waits. This may feel smoother than forcing the user to pause after every action. It also creates a new design responsibility: the agent must distinguish a task in progress from a completed one, so fluent speech does not imply success before confirmation.
Extended Thinking is intended for requests needing more steps. Google says it can reason in the background and narrate progress without breaking the conversational flow. That can help users understand why a complex task is taking time, but progress narration must remain accurate. A reassuring phrase is not evidence that the promised result has been checked.
Benchmarks need their context
Google reports that Extended Thinking leads several speech-to-speech and agentic-task benchmarks, including an 82.6 score on the Artificial Analysis Speech to Speech Quality Index and a 68.6 per cent result on τ-Voice. These figures are useful comparisons under the stated test conditions, but they should not be read as success rates for every deployment. Voice tasks vary in noise, language, interruptions, tool latency and the consequences of a wrong action.
The announcement also says Gemini 3.8 Live placed second in a user-preference arena. Preference can capture how natural a conversation feels, while a task benchmark asks whether the agent completed defined work. Both dimensions matter. A pleasant voice that makes incorrect bookings is not a good service; an accurate system that routinely interrupts or loses context can also fail users.
An organisation assessing the models should run its own representative calls, measuring task completion, correction time and escalation to a person. Testing only a scripted demonstration would miss the rough edges that appear in real conversations.
Where it is available
Google’s Gemini API release notes label the two audio-to-audio models generally available on 15 September. Developers can use the Live API, subject to the service terms and supported configuration. The broader article discusses use across the Gemini app, Workspace and Search, but that should not be collapsed into a promise that every feature is present for every account today.
A developer choosing between the models should start with the task. A low-latency information exchange may favour standard Live; a multi-step request may justify Extended Thinking. The extra reasoning is useful only when its contribution outweighs added complexity or cost. Google’s pricing and model documentation should be checked at implementation time because those details can change.
The models also differ from a conventional transcription pipeline. Audio-to-audio handling can preserve conversational cues, but teams may still need a separate transcript for review, accessibility or records. How those records are created and retained belongs in the application design.
A careful production trial
A bounded pilot might ask the assistant to answer questions about an approved knowledge base, then retrieve a non-sensitive account detail with explicit permission. Reviewers should test interruptions, language switches, a delayed tool response and a failed tool response. They should confirm that the model says when it does not know and does not claim an action happened before the system confirms it.
Where visual input is involved, privacy deserves equal attention. A camera may show people, documents or locations beyond the object of interest. The application should make capture and sharing understandable to the user rather than relying on the model to infer consent.
Google’s release is significant because it treats voice, vision and asynchronous actions as parts of one agent experience. The opportunity is a more natural route to useful work. The limit is equally clear: benchmark leadership and smooth demos still need to translate into reliable, consent-aware behaviour on the specific tasks people actually ask the system to perform.