Google

AI · Voice AI · Developer Tools

Launch

Google launches Gemini 3.8 Live voice models with background reasoning

The models keep talking while fetching data or calling tools, closing the awkward silent gap that has dogged voice AI agents.

• Google DeepMind released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026, two new audio-to-audio models for real-time voice agents via the Gemini API and Google AI Studio.
• Gemini 3.8 Live is built for fast, cost-efficient conversations with visual grounding, while Extended Thinking targets high-complexity, multi-step reasoning tasks.
• Extended Thinking can reason and call tools asynchronously in the background while narrating progress verbally, so a completed spoken turn no longer means the model has gone idle.
• Both models handle text, image, audio and video input, support 97+ languages with mid-conversation switching, and cap at 131,072 input tokens and 65,536 output tokens.
• Extended Thinking topped Artificial Analysis' Speech-to-Speech Quality Index with an 82.6 score, edging out OpenAI's GPT-Live-1.
• Google prices the models at $0.005 per minute of audio input and $0.018 per minute of output, versus OpenAI's roughly $0.05 per minute for GPT-Live-1 — about $1.38 versus $3+ for an hour of conversation.
• 3.8 Live is already rolling out to power Search Live in the Google app, alongside Google Workspace and the Gemini app, letting users switch languages mid-conversation and follow web links from spoken answers.
• By making voice agents narrate their own background work instead of going silent, Google is targeting the core UX flaw that has held back voice AI for tasks like support, tutoring and multi-step automation.