Google Releases Gemini 3.8 Live and Extended Thinking Real-Time Conversation Models

On September 15 local time, Google officially announced the release of two new products: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Google stated that these are its most advanced real-time conversation models to date, aiming to make AI conversations more intuitive and intelligent through significant upgrades in intelligence and parallel reasoning. These two models are optimized for different application scenarios: Gemini 3.8 Live focuses on scalable deployment and cost efficiency, while Gemini 3.8 Live Extended Thinking is dedicated to high-complexity tasks and enhanced multi-step reasoning capabilities.

In terms of performance, Gemini 3.8 Live Extended Thinking achieved leading results in multiple benchmark tests. It ranked first overall in the Artificial Analysis Voice-to-Voice Quality Index with a score of 82.6%, and achieved a task completion rate of 68.6% in the τ-Voice agent task. Meanwhile, the model scored 97.7% in the Big Bench Audio test, significantly raising the intelligence level of complex workflows while maintaining competitive pricing.

Gemini 3.8 Live has received high recognition for user experience and cost-effectiveness, ranking second in the Speech Agent Arena. The model features excellent visual understanding capabilities, capable of processing visual inputs in near-real-time, and automatically detecting and switching among 97 supported languages during conversations. Additionally, it can execute tool and API calls in the background while keeping the conversation smooth and uninterrupted, successfully balancing accuracy and conversational quality, thereby pushing the Pareto frontier of complex workflows.

For deep reasoning needs, Gemini 3.8 Live Extended Thinking achieves synchronous reasoning and speaking. Without interrupting the fluency of the conversation, it responds with natural language prompts like 'Let me confirm...' and can explain the progress of multi-step backend tasks in real-time. This design allows the model to provide accurate logical deduction while maintaining a natural interactive experience when handling high-complexity tasks.

In terms of ecosystem building, developers can easily build high-performance voice-driven interfaces via the Gemini Live API on platforms such as Agora, Fishjam, LiveKit, Pipecat, Vercel, Vision Agents, etc. Google is also collaborating with enterprises such as Salesforce, Genspark, and Lumeris to jointly promote the development of the relevant ecosystem, aiming to apply advanced real-time conversation capabilities more widely to various commercial scenarios.

Security and compliance are important components of this release. Google has added SynthID invisible watermarks to all AI-generated audio, embedded directly into the audio output. This technology aims to help detect and prevent the spread of misinformation, ensure the traceability of AI-generated content, and provide a safer AI interaction environment for users and enterprises.

Currently, Gemini 3.8 Live has begun rolling out. Developers can use it in the Gemini API and Google AI Studio; enterprise users can access the early preview in Gemini Enterprise, which will soon land on Gemini Enterprise for Customer Experience; all users can experience it in Search Live. The popularization of this model marks the implementation of real-time conversation technology in more scenarios.

Gemini 3.8 Live Extended Thinking has also started rolling out. Developers can similarly use it in the Gemini API and Google AI Studio; enterprise users can access the early preview in Gemini Enterprise, which will soon land on Gemini Enterprise for Customer Experience and Google Workspace enterprise customers; general users can experience it in Gemini Live, Google AI Pro and Ultra subscribers can use it in Docs within Google Workspace, and all Google AI subscribers can experience it in Gmail and Keep.