SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Gemini 3.8 Live - Google's native real-time voice conversation model

Author: AI小集 · Original source: AI工具集

What is Gemini 3.8 Live

Gemini 3.8 Live is a series of real-time voice conversation models launched by Google, including Gemini 3.8 Live, which is aimed at deployment at scale and balances conversational intelligence, visual understanding, and cost-effectiveness; Gemini 3.8 Live Extended Thinking strengthens multi-step reasoning and can broadcast progress by voice while reasoning. The entire model series adopts a native speech-to-speech architecture, supports automatic switching among 97 languages and asynchronous background tool calls, tops the Artificial Analysis voice quality index, and generates audio with SynthID watermarks. It has been rolled out to developers, enterprises, and ordinary users.

Gemini 3.8 Live

Main features of Gemini 3.8 Live

  • Real-time voice conversation: Conducts conversations in a native speech-to-speech manner with near-real-time responses, ranking second in user preference on the Speech Agent Arena.
  • Asynchronous tool calls: Executes API and tool calls in the background while continuously broadcasting voice replies to the user, without interrupting the conversation.
  • Visual understanding: Processes camera footage in near real time, providing visual context for the conversation.
  • Multilingual switching: Supports 97 languages, automatically detecting and seamlessly switching languages during a conversation.
  • High alphanumeric accuracy: Accurately parses technical string data such as confirmation codes and claim numbers.
  • Incremental content updates: Seamlessly integrates real-time audio with structured data and returns context-aware responses.

Technical principles of Gemini 3.8 Live

  • End-to-end speech-to-speech architecture: The model abandons the traditional cascaded pipeline of "speech recognition → text large model → speech synthesis" and directly processes input and output in an audio-to-audio manner, significantly reducing latency and preserving acoustic details such as tone and speaking rate. It is a continuation and upgrade of the previous Gemini 3.1 Flash Live architecture route.
  • Parallel reasoning and "thinking while speaking": The Extended Thinking variant adopts an asynchronous reasoning mechanism, continuing to stream voice output while performing multi-step thinking in the background. Through early verbal cues such as "Let me confirm that...", it first confirms the request, then broadcasts task progress in real time, achieving parallel reasoning and speaking without freezing the conversation.
  • Configurable thinking budget: Extended Thinking supports developers in configuring thinking intensity levels as needed, investing more background reasoning compute in complex tasks and reducing latency and cost in simple conversations. Official pricing is $0.005 per minute for audio input and $0.018 per minute for output.
  • Full-chain content watermarking: All generated audio embeds an invisible SynthID watermark directly woven into the audio signal itself. It can still be detected even after editing or compression, and is used to prevent voice forgery and the spread of false information.

How to use Gemini 3.8 Live

Gemini 3.8 Live：
  • Developers: Can use it in the Gemini API and Google AI Studio .
  • Enterprise users：Gemini Enterprise is available in private preview.
  • All users: Can experience it in Search Live.
Gemini 3.8 Live Extended Thinking：
  • Developers: Can use it in the Gemini API and Google AI Studio.
  • Enterprise users: Available in private preview in Gemini Enterprise, and will soon be available to Gemini Enterprise for Customer Experience and Google Workspace enterprise customers.
  • All users: Available in Gemini Live; Google AI Pro and Ultra subscribers can use it in Docs in Workspace, and all Google AI subscribers can use it in Gmail and Keep.

Core advantages of Gemini 3.8 Live

  • End-to-end speech architecture: Native speech-to-speech processing delivers lower latency than cascaded solutions and better preserves acoustic details such as intonation.
  • Outstanding cost-effectiveness: Audio input is $0.005/minute and output is $0.018/minute, making it suitable for large-scale deployment.
  • Asynchronous tool calling: Executes API and function calls in the background while continuing voice replies, so multi-step tasks do not interrupt the conversation.
  • Real-time visual understanding: Can process camera footage in near real time, enabling voice agents that can "listen and see."
  • Automatic switching across 97 languages: Automatically detects and seamlessly switches languages during conversations, covering users worldwide.
  • Parallel reasoning capability: The Extended Thinking variant reasons while speaking, using verbal cues such as "let me confirm that..." to keep the conversation flowing.
  • Benchmark leader: Extended Thinking tops the Artificial Analysis Speech Quality Index (82.6), with a Big Bench Audio score of 97.7%.

Comparison of Gemini 3.8 Live with similar competitors

Comparison item Gemini 3.8 Live OpenAI gpt-realtime-2.1
Model architecture Native speech-to-speech, a streamlined alternative to cascaded solutions Native speech-to-speech, with GPT-5-level reasoning built into the audio loop
Pricing model Billed per minute Billed per token
Audio input cost $0.005/minute $32 per million tokens (about $0.019/minute)
Audio output cost $0.018/minute $64 per million tokens (about $0.077/minute)
Typical session cost About $0.005–0.018/minute In practice about $0.06–0.11/minute; for long conversations without caching, up to $0.18–0.46/minute
Language support 97 languages, with automatic detection and switching during conversations Real-time speech translation requires a separate model, supporting 70+ input languages to 13 output languages
Visual understanding Supported, with near-real-time processing of camera footage (up to about 1 frame/second) Does not provide equivalent real-time visual voice conversation capabilities
Deep reasoning The Extended Thinking variant speaks while reasoning, with configurable thinking levels Has a built-in reasoning loop, but no "think aloud" progress announcement mechanism
Benchmark results No. 1 on the AA Speech Quality Index (82.6), Big Bench Audio 97.7% Big Bench Audio improved 15.2% over the previous generation
Content safety All generated audio is embedded with the invisible SynthID watermark No equivalent audio watermarking mechanism

Application scenarios of Gemini 3.8 Live

  • Employee onboarding guidance: Through real-time visual understanding via camera, new employees can ask questions by voice while operating, and the AI immediately answers process questions.
  • Intelligent customer service and banking operations: Combine asynchronous function calls to complete multi-step business processing, with real-time progress announcements during background processing, corresponding to the τ-Voice-banking benchmark scenario.
  • Travel and Booking Assistant: After the user makes a multi-condition request by voice, the AI asynchronously calls ticketing and hotel APIs in the background, keeping the natural conversation uninterrupted while querying.
  • Programming and Design Collaboration: A developer describes requirements by voice based on a sketch, and Extended Thinking generates usable React component code while reasoning.
  • Real-time Troubleshooting: The user describes the problem by voice in Search Live, and the AI combines visual images to provide step-by-step voice guidance for troubleshooting and repair.