Unveiling Gemini 3.8 Live & 3.8 Live Extended Thinking

Published: September 15, 2026
Authors: Tom Ouyang (Principal Engineer) and Malini Jaganathan (Member of Technical Staff), representing the Gemini Audio Team.

Google has officially introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, marking a significant leap in the evolution of live dialogue models. These updates focus on enhancing parallel reasoning and intelligence, making voice-based collaboration more intuitive and capable of executing intricate tasks.

Header Image

🎙️ A New Era of Voice Interaction

The goal of these models is to transform AI from a tool into a true conversation partner. By integrating near real-time reasoning, Google is making voice agents feel more natural and fluid.

~~Old AI felt robotic and linear~~ $\rightarrow$ New AI feels intuitive and collaborative.

Key Capabilities:

Fluidity: Handles interruptions and switches languages seamlessly.
Transparency: The AI can articulate its internal thought process while working.
Multitasking: Executes background tasks and manages tools without breaking the conversational flow.
Contextual Awareness: Integrates real-time visual grounding and language support.

🛠️ Model Comparison: Which one to use?

Depending on the use case—whether it's a lightweight consumer app or a heavy-duty enterprise solution—Google offers two distinct paths:

Feature	`Gemini 3.8 Live`	`Gemini 3.8 Live Extended Thinking`
Primary Focus	Scale & Cost-Efficiency	High-Complexity Tasks
Core Strength	Fluid dialogue & Visual grounding	Multi-step reasoning & Intelligence
Ideal Use Case	Fast, responsive voice interactions	Enterprise-grade task completion
Performance	Optimized for speed/cost	Optimized for depth/accuracy

📈 Technical Benchmarks & Performance

The Gemini 3.8 Live Extended Thinking model has set new industry standards for speech-to-speech quality and agentic capabilities.

Quantitative Results:

Speech to Speech Quality Index: Ranked #1 overall with a score of $\text{82.6}$ .
$\tau$ -Voice (Agentic Task Completion): Achieved $\text{68.6\%}$ .
Sierra’s $\tau$ -Voice-banking Benchmark: Achieved $\text{35.1\%}$ .

Performance Visuals:

Speech to Speech Index Agentic Performance Sierra Banking Benchmark Cost per Hour EVA Bench

⚙️ How it Works: The Reasoning Flow

The "Extended Thinking" process allows the model to process complex requests in the background while maintaining a front-facing conversation.

For developers, implementing these models via the API might look like this conceptually:

{
  "model": "gemini-3.8-live-extended-thinking",
  "input": "voice_stream",
  "capabilities": ["visual_grounding", "parallel_reasoning"],
  "output_mode": "real_time_audio"
}

🚀 Availability & Deployment

These models are designed to be production-ready for both individual users and large-scale enterprises.

Gemini App (Consumer access)
Google Workspace (Productivity integration)
Google Search (Enhanced voice queries)
Gemini API (Developer and Enterprise access)

💬 Industry Feedback

Leading organizations are already integrating these capabilities into their workflows:

Salesforce

11Sight

Equal AI

ServiceNow