Real-Time Audio Models That Think While Speaking: Google Releases Gemini 3.8 Live and Extended Thinking, Pushing Real-Time Voice Toward Multi-Step Reasoning

The Google Gemini Audio Team has released two new real-time voice models: Gemini 3.8 Live for scale and cost efficiency, and Gemini 3.8 Live Extended Thinking for complex tasks and multi-step reasoning. This article covers pricing, benchmark results, seven demonstrations, and setup instructions.

Quick Take

• Google has introduced two real-time voice conversational models: Gemini 3.8 Live, focusing on scale and cost efficiency, and Gemini 3.8 Live Extended Thinking, specializing in complex tasks and multi-step reasoning.
• The core breakthrough is a parallel "thinking while speaking" reasoning mechanism: when performing background tool calls or deep planning, the model can provide early verbal confirmations and maintain conversational continuity through real-time progress narration.
• Setting new records across multiple speech and agent benchmarks, Extended Thinking achieved first place on the Speech-to-Speech Index and leads in task completion rates across τ-Voice and Sierra banking evaluations.
• The seven official demos cover onboarding, chess, sketch-to-React generation, restaurant booking, brand design, Workspace collaboration, and Search Live plumbing repair.
• One minute of listening costs about $0.005. If the model speaks for 30 seconds, the total is about $0.014; if it speaks for the full minute, the total is about $0.023. All three Live models have the same price, and the free tier costs $0.
• Gemini 3.8 Live is open to developers via the Gemini API and Google AI Studio; Extended Thinking has entered private preview for enterprises and is rolling out to Gemini Live and select Workspace subscription tiers; audio generated by Google AI products embeds imperceptible SynthID watermarks.

Division of Labor Between Dual Models: Lightweight Scale Deployment vs. High-Complexity Deep Reasoning

When processing complex tasks in real-time voice, a system must respond as quickly as possible while maintaining conversational continuity during reasoning and tool calls. Rather than using a single model to cover all scenarios, Google split scale and cost efficiency from high-complexity reasoning across two versions.

The base version, Gemini 3.8 Live, focuses on cost efficiency for large-scale deployments. Combining conversational intelligence with near-real-time visual understanding and multilingual streaming transitions, it targets high-concurrency, low-latency everyday assistant and retrieval services. The model supports automatic recognition and smooth transitions across 97 languages during conversations, accurately parses alphanumeric strings according to official claims, and supports incremental content updates and asynchronous function calling—all while continuing the conversation as tools and API calls execute in the background.

Targeting demanding workflows, Gemini 3.8 Live Extended Thinking focuses on enterprise-grade complex task execution. It significantly strengthens multi-step logical reasoning and planning capabilities, primarily aimed at production environments requiring deep thinking such as customer service, specialized troubleshooting, and multi-system coordination. Together, the two models form a complementary hierarchy ranging from low-cost interaction to underlying complex business workflows. Within the same audio product portfolio, Gemini 3.5 Transcribe is a dedicated speech-to-text model supporting over 85 languages, with an officially reported average word error rate (WER) of 4.0% in streaming and 2.6% in non-streaming mode; it is not a replacement for 3.8 Live, but rather a pure transcription component operating at a different architectural layer.

Pricing: What Does One Minute Cost?

All three models have the same price. For one minute:

Audio input costs $0.005 per minute and audio output costs $0.018 per minute. Thirty seconds of model speech therefore costs $0.009, for a total of about $0.014 including listening.

The free tier costs $0. Optional transcription and Google Search Grounding charges are not included above.

Thinking While Speaking and Background Execution: Minimizing Awkward Silences in Complex Tasks

The most disruptive experience in real-time voice interaction is prolonged system silence after posing a complex request. The pivotal design of Extended Thinking is reasoning and speaking simultaneously: verbally acknowledging receipt of the request first, and then narrating progress aloud as multi-step tasks advance.

This real-time coordination is underpinned by three core interaction mechanisms:

The first is early verbal confirmation. When a user issues a request requiring retrieval or computation, the model first provides natural spoken feedback (such as "Let me see" or "I am looking up that record"), maintaining continuous communication through early acknowledgment and preventing the anxiety-inducing silence typical of traditional voice interfaces.

The second is real-time progress narration. When tackling background tasks that require multi-endpoint calls or staged execution, the model does not wait for all results to return before speaking; instead, it synchronizes ongoing progress with the user through spoken narration as reasoning and task workflows unfold, sustaining an unbroken dialogue.

The third is parallel background execution and dialogue. The model continues conversing while executing tools and API calls in the background, preserving bidirectional communication throughout background task processing.

Benchmark Results

Google reported the following results:

• Extended Thinking: 82.6% on Speech to Speech Index, first overall; 68.6% on τ-Voice; 35.1% on Sierra τ³-Banking; and 97.7% on Big Bench Audio.
• Base 3.8 Live: 76.0% on Speech to Speech Index; 30.1% on τ-Voice; the lowest Big Bench Audio hourly cost in the table at $0.84; and second place in Speech Agent Arena user preference.
• Versus 3.1 High: Extended Thinking leads by 11.1, 30.9, and 23.8 percentage points on Speech to Speech Index, τ-Voice, and Sierra τ³-Banking. Base 3.8 Live is 4.5 points higher on Speech to Speech Index and 7.6 points lower on τ-Voice.
• EVA-Bench: The 3.8 models sit on the Pareto frontier for accuracy and conversational experience.

The blue dashed line marks the Pareto frontier reported by Google.

What the Seven Official Demos Show

Google published seven demonstration videos covering visual perception, code generation, conversational interruptions, and cross-application workflows.

Onboarding Assistant: The model guides a new employee named David through a dashboard, identifies the “Personal Information” icon, confirms the Health & Benefits page, and explains HMO. When asked for answers to an information-security quiz, it refuses and tells the employee to answer independently.

Real-Time Chess Match: The model combines visual board recognition with spoken moves. The user plays White with d4, e4, e5, Qf3, Bc4, and Qxf7#; Gemini plays Black and is checkmated.

Real-Time Sketch to React Component: The user draws a phone interface on paper and keeps requesting changes by voice. The model adds placeholder product images, copy, a bottom drawer, and an icon-based action list, then switches the design to glassmorphism. Koray Kavukcuoglu's 71-second X post contains the same clip; this article keeps the translated version with Chinese-English subtitles.

Restaurant Booking and Interruption Handling: The user books a table for eight next Saturday and asks about bringing a cake, then interrupts with a food image and asks whether it is vegetarian. The model identifies Venezuelan cheese sticks (Tequeños), answers the question, returns to the booking flow, collects contact details, and completes the reservation.

Ramen Shop Brand and Marketing Kit: The user asks for a mix of Japanese tradition and Bushwick street character with minimalist red accents. The model creates a mood board, revises an urban ink-wash scene by voice, and expands the work into brand guidelines and a logo system.

Deep Google Workspace Collaboration: In Gmail, the model extracts a preschool checklist from email; in Keep, it organizes painting steps and materials; in Docs, it drafts a neighborhood party plan and retrieves booked hotel and travel dates.

Search Live Physical Plumbing Troubleshooting: The user points a phone camera at household plumbing. The model identifies the blue-handled cold-water main valve, reads a caliper measurement as nominal three-quarter-inch copper pipe, explains nominal versus actual diameter, and warns about the skill and leak risk involved in cutting the main line.

How to Use: From Direct Evaluation to SDK Integration

Whether for no-code hands-on evaluation or engineering development, Google provides onboarding paths spanning web consoles to code SDKs:

1. Regular Users: Direct Zero-Configuration Experience

• AI Studio Web Console: Access the Live console directly in Google AI Studio, select gemini-3.8-live or gemini-3.8-live-extended-thinking from the model dropdown, and engage in full-duplex voice and visual interactions using your microphone and camera.
• Consumer and End-User Surfaces: General users can experience 3.8 Live in Search Live; Gemini Live and features in Gmail, Docs, and Keep are rolling out in phases across consumer and enterprise subscription tiers.

2. Developers: Minimal Runnable Python Integration Example

Developers can establish bidirectional WebSocket streaming sessions using the latest official google-genai SDK.

Prerequisites:

pip install -U google-genai
export GEMINI_API_KEY="your-api-key-here"

Minimal Connection Code:

import asyncio
import os
from google import genai

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
model = "gemini-3.8-live"
config = {"response_modalities": ["AUDIO"]}

async def main():
    async with client.aio.live.connect(model=model, config=config) as session:
        print("Session started")

if __name__ == "__main__":
    asyncio.run(main())

3. Key Selection and Migration Considerations

• Model Selection: gemini-3.8-live is suitable for low-latency everyday conversation, direct commands, and lightweight, fast-responding tools; gemini-3.8-live-extended-thinking is suited for workflows demanding multi-step planning, diagnostics, longer-running tools, and scenarios where the model must verbally bridge wait times while reasoning in the background.
• Migrating from 3.1 to Base 3.8: Update the model identifier to gemini-3.8-live, and ensure removal of legacy thinking_level and thinking_config parameters (base 3.8 does not support any thinking configurations); the turnComplete lifecycle remains consistent.
• Extended Thinking Implementation Standards:
  • The model ID must be specified as gemini-3.8-live-extended-thinking.
  • Function and tool definitions must explicitly declare behavior: NON_BLOCKING.
  • Task Lifecycle Evaluation: The completion of a single spoken audio turn merely marks the end of a voice turn (signaled by turnComplete: true) and does not signify overall task completion; the overarching state must be determined via interaction_status. The model is only truly idle when the status transitions to IDLE; an IN_PROGRESS state indicates the model is still reasoning in the background, executing tools, or preparing further responses.
  • Thinking Levels: thinking_level supports low, medium, and high; minimal is not supported.
• Client-Side Security Architecture: When connecting to the Live API directly from browsers or mobile clients, never hardcode long-lived API keys on the client; short-lived ephemeral tokens must be issued by a backend authentication server, keeping long-lived credentials secured server-side.

Developer Ecosystem and Streaming Infrastructure: Lowering the Barrier to Real-Time Full-Duplex

Building a highly available, low-latency real-time voice conversational system involves model inference as merely the tip of the iceberg. For the vast majority of application developers, managing complex real-time media streaming infrastructure is exceptionally burdensome.

Google has adopted an ecosystem partnership strategy. Alongside the opening of the Gemini Live API, developer platforms and communications cloud providers including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents have joined as partners. These partner platforms manage complex real-time streaming media infrastructure, enabling engineering teams to build custom voice agents using their respective SDKs.

Google also listed partners including Salesforce, Genspark, and Lumeris, which highlighted latency, conversational fluency, and tool-calling capabilities.

SynthID Watermarking and Rollout Scope

To ensure transparency and provenance, audio generated across all Google AI products embeds imperceptible SynthID watermarks, keeping AI audio reliably detectable.

The rollout for both models is progressing in phases:

Gemini 3.8 Live is available to developers via the Gemini API and Google AI Studio; for enterprises, it is in private preview on Gemini Enterprise and slated for Customer Experience; and consumer users can access it in Search Live.

Gemini 3.8 Live Extended Thinking is currently in private preview for enterprise customers, followed by Customer Experience and Workspace commercial plans; consumer access includes Gemini Live, Workspace Docs for Pro and Ultra subscribers, and Gmail and Keep for all Google AI subscribers.

Summary

Gemini 3.8 Live keeps the same price as 3.1 and is the choice for low-latency, high-volume voice applications. Extended Thinking combines multi-step reasoning, background tool calls, and continuous spoken feedback in one session, making it better suited to complex customer service, troubleshooting, and cross-system workflows. The choice is simple: use 3.8 Live for cost and speed, and Extended Thinking for task completion and complex reasoning.