Google has released two new Gemini voice models designed to make conversations with artificial intelligence faster, more natural and better suited to complex tasks.
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking became generally available through Google’s Live API on September 15, with Google describing them as its latest audio-to-audio models for real-time voice applications.
Table of Contents
What the Gemini update changes
The standard Gemini 3.8 Live model is intended for low-latency voice interactions where users want quick responses without a noticeable reasoning pause. Google says it supports interleaved reasoning, asynchronous function calling and full session content updates for developers building live voice experiences.
Gemini 3.8 Live Extended Thinking is aimed at more demanding workflows. It can reason in the background while continuing to speak with the user, allowing it to handle multi-step tasks without ending or repeatedly interrupting the conversation.
That approach is designed for voice agents that need to do more than answer questions, such as coordinating bookings, completing workplace tasks or working through a complicated troubleshooting process.
How Gemini 3.8 Live works
Google says the models can process visual information in near real time, giving conversations additional context beyond spoken words. The company also says the system can detect and switch between 97 supported languages during a conversation.
Another major feature is background tool use. Gemini can make tool or application programming interface calls while the conversation continues, rather than forcing the user to wait in silence for each operation to finish.
For example, a voice agent could acknowledge a request, begin checking information or completing a task, and then return with the result while maintaining a conversational exchange. Google demonstrated use cases involving employee onboarding, live troubleshooting, business planning and multi-step bookings.
What the new models mean for developers
For developers, the update is less about a new chatbot interface and more about infrastructure for building voice-based software. The Live API provides access to the two models, with Gemini 3.8 Live positioned as the general-purpose option and Extended Thinking intended for applications where deeper reasoning is more important than the lowest possible delay.
That distinction could allow companies to choose between speed and task complexity instead of using one model for every voice interaction.
Google is also presenting the models as suitable for production use by businesses. The company says Gemini 3.8 Live Extended Thinking placed first on Artificial Analysis’s Speech to Speech Quality Index, with a score of 82.6, and reported results on agentic voice and audio reasoning benchmarks. Those figures are company-cited performance claims and should not be treated as an independent assessment of every real-world use case.
Where users may encounter Gemini 3.8 Live
Google says the new Live models are being used across the Gemini app, Google Workspace and Search, including voice-driven experiences that can help users manage tasks, work through documents or receive step-by-step assistance.
The release follows a series of rapid Gemini updates in September. Google made Gemini 3.8 Flash generally available on September 2 for software engineering, autonomous agents and enterprise workflows, before releasing the two Live models later in the month.
Google also launched a Gemini desktop app for Windows on September 11. The app can be summoned with the Alt + Space shortcut and can work alongside active applications, connect with Gmail and Drive for certain tasks, and generate images through Google’s Nano Banana system.
What happens next
The immediate focus will be on how developers use the new models to create voice agents that can act, not merely respond. Background reasoning, visual inputs and asynchronous tool calls could make voice interfaces more useful for customer service, workplace software, education and accessibility applications.
The same capabilities also raise practical questions about privacy, reliability and user control. Systems that can listen, interpret visual context and call external tools need clear permissions and safeguards, particularly when they are connected to workplace accounts or personal information.
For now, Google’s release marks another step in its effort to move Gemini from a question-and-answer assistant toward a continuous voice interface that can understand context, reason through tasks and take action during a conversation.
Frequently Asked Questions
What is Gemini 3.8 Live?
Gemini 3.8 Live is Google’s low-latency audio-to-audio model for real-time conversations and voice agents.
What is Gemini 3.8 Live Extended Thinking?
It is a higher-reasoning model that can work through complex, multi-step tasks in the background while continuing the conversation.
When were the new Gemini Live models released?
Google listed both models as generally available through its Live API on September 15, 2026.
What can Gemini 3.8 Live do?
The models can support spoken conversations, process visual context, switch between supported languages and make tool or API calls during an interaction.
Are Gemini 3.8 Live and Extended Thinking the same model?
No. Gemini 3.8 Live is optimized for speed and fluid dialogue, while Extended Thinking is designed for more complex reasoning and workflows.
Fact-Checked: Key release dates, model descriptions and availability details were checked against Google’s official announcement, Gemini API release notes and Google Workspace documentation.
Disclaimer: Availability and features may vary by account, region, product and developer access level.