Explainers / tech / № 1363

Gemini 3.8 Live lets voice agents reason without going quiet

Google has released two speech-to-speech models for its Gemini API and AI Studio, adding background tool calls and configurable reasoning to live conversations.

· 5 min read

✓ Verified Source Google DeepMind's developer announcement, independently cross-checked against TechRepublic, ExtremeTech, Unite.AI and 9to5Google coverage; leaderboard, language-count, pricing and capability figures remain primarily Google-reported. ⚑ Voice AI

The 60-second version

Google's Gemini 3.8 Live models let voice agents speak while background tools and configurable reasoning run.

Key points

  • Asynchronous function calling is meant to prevent tool waits from turning into dead air.
  • Google highlights visual context, alphanumeric precision and coverage for more than 97 languages.
  • The Extended Thinking leaderboard result, prices and capability figures are vendor-reported.
  • The real test is completed-task reliability: correct tool choice, clear reporting and recovery from failure.

Verdict. A meaningful voice-agent architecture update, but a smoother conversation is not evidence that the agent will take the right action.

What shippedA voice agent that can keep the floor while it works

Google has released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking through the Gemini API and Google AI Studio. Both are native speech-to-speech models aimed at developers building real-time, voice-first products.

The central change is asynchronous function calling: an agent can continue streaming audio while an API or tool call runs in the background. In principle, a scheduling assistant can acknowledge a request naturally while checking a calendar, instead of going silent until the action finishes.

97+languages Google says the Live models cover
$0.005per minute for listed audio input
$0.018per minute for listed audio output

Why it mattersThe pause is part of the product

Voice agents are judged by timing as much as by words. A long silence makes a system feel stuck; speaking too early can make it claim an action before the tool has returned. Background calls give the agent a way to manage that tension, although they do not guarantee that the eventual action is correct.

More than a microphone

  • Visual context lets a live conversation refer to what the user is showing the agent.
  • Google highlights better handling of confirmation codes, claim numbers and other alphanumeric data.
  • The Extended Thinking model adds configurable reasoning for complex, multi-step requests.
  • Google lists integration partners including LiveKit, Agora, LangChain and Vercel.

Read the claim carefullyA smoother turn is not the same as a safer action

Google says Gemini 3.8 Live Extended Thinking ranked first on Artificial Analysis' speech-to-speech leaderboard. That result describes one evaluation setup. It does not establish universal accuracy, low latency, safe tool use or robustness in noisy environments.

The bottom lineThe architecture is the interesting part

Gemini 3.8 Live targets a real weakness in voice interfaces: the awkward gap between conversation and action. If developers can make background work feel natural without hiding uncertainty, the release could make voice agents more usable. The launch itself proves a new API shape and a new set of promises; it does not yet prove dependable autonomy.

The real upgrade is not that the agent talks more. It is that the conversation can continue while the system does the work.

Primary sourcesGoogle DeepMind: Build real-time voice applications with Gemini 3.8 Live and 3.5 Transcribe·Google News RSS: independent coverage of Gemini 3.8 Live·TechRepublic: Google Launches Gemini 3.8 Live Models That Can Reason While They Talk·ExtremeTech: Google Launches New Voice AI Models for Building Real-Time Conversational Apps·Google Gemini API Live API documentation