Google has released Gemini 3.8 Live and 3.8 Live Extended Thinking, new audio models that let AI agents reason deeply while maintaining natural conversation flow, without the awkward pauses users typically experience. Launched on September 15, these models represent a fundamental shift in how conversational AI works, separating the speed of responding from the complexity of thinking.

What Makes These Audio Models Different From Previous AI Assistants?

The key innovation lies in how these models handle two different types of delay. Interactive latency is the time between when you speak and the AI starts responding. Reasoning latency is the total time the AI needs to think through a complex problem. Previous voice AI models forced users to wait for both to finish before getting any response, creating awkward silences.

Gemini 3.8 Live decouples these two processes. The model can start streaming audio responses to you immediately while simultaneously performing background reasoning, making API calls, and executing tasks in the background. This means the conversation never stops, even when the AI is doing heavy computational work.

"The model is designed to do background reasoning. It's doing it during the conversation so you can keep the conversation alive while it's reasoning," said Sid Nag, founder of Tekonyx.

Sid Nag, Founder at Tekonyx

The Extended Thinking version adds configurable reasoning, allowing users to toggle the AI's internal thought process on or off depending on the task complexity. For straightforward questions, you get instant answers. For multi-step problems, the AI can narrate its reasoning while working through the solution.

How Do These Models Perform Compared to Competitors?

Google's new speech-to-speech models rank first on Artificial Analysis' Speech-to-Speech leaderboard, a widely used benchmark for comparing voice AI systems. The company also released Gemini 3.5 Transcribe, a dedicated speech-to-text model that achieves a 4.0% word error rate in streaming mode and 2.6% in non-streaming mode, meaning it accurately transcribes speech across 85+ languages with minimal mistakes.

The models support more than 97 languages and can handle code-switching, where speakers mix languages mid-sentence without any manual configuration. They can also parse alphanumeric data like confirmation codes and claim numbers with high precision, a critical feature for customer service applications.

"An AI agent doesn't necessarily have to choose between being fast and being thoughtful. It can actually maintain a real-time interaction," noted Sid Nag.

Sid Nag, Founder at Tekonyx

How to Build Real-Time Voice Applications With These Models

• Asynchronous Function Calling: Execute API calls and tool integrations in the background while continuing to stream audio responses to users, eliminating the need to pause conversations for data lookups or system integrations.
• Visual Context Integration: Ground conversations in live visual inputs so agents can understand both what users say and what they see, enabling multimodal applications that combine voice, text, and video.
• Custom Vocabulary Biasing: Steer speech recognition toward domain-specific terms, company names, and technical jargon by passing a custom vocabulary list of up to 1,000 terms for specialized industries.
• Incremental Content Updates: Seamlessly merge real-time audio responses with structured data like search results or database queries to deliver context-aware answers without interrupting the conversation flow.

Developers can access these models through the Gemini API and Google AI Studio, with pricing set at $0.005 per minute for audio input and $0.018 per minute for audio output, making them cost-effective for scaling voice applications. Integration partners including Agora, Fishjam, LangChain, LiveKit, Pipecat, and Vercel handle the media streaming infrastructure, so developers don't need to build that layer themselves.

What Real-World Problems Can These Models Solve?

Enterprise applications include customer support chatbots that can look up account information while answering questions, sales processes where agents need to access pricing and inventory data mid-conversation, and multimodal agentic applications that combine text, voice, and video inputs. A customer calling a support line could describe a problem, and the AI agent could simultaneously search a knowledge base, check system logs, and propose solutions without asking the customer to wait or repeat information.

The models also enable real-time translation across more than 70 languages, live transcription for accessibility and compliance, and voice-first interfaces for hands-free applications like smart home control or in-vehicle assistants.

"The way it's architected, you can customize this to behave in a lot of different ways," said Bradley Shimmin, an analyst at Futurum Group. "You can inject context into the conversation without interrupting what it's doing."

Bradley Shimmin, Analyst at Futurum Group

Google's approach appears to be catching up to competitors like Apple, which already offers real-time transcription in productivity apps such as Notes and Voice Memos. However, Shimmin noted that Google's architecture takes the concept further by enabling true conversational AI rather than simple prompt-response interactions.

The release of Gemini 3.8 Live comes two weeks after Google introduced the Gemini 3.8 Flash and 3.8 Cyber foundation models, signaling the company's broader push into multimodal AI systems that can process and reason across audio, visual, and text inputs simultaneously.