Introduction

Producers live for the moment a take stays in the pocket while the rest of the rack does heavy work in the background. On 15 September 2026, Google dropped that same energy into voice AI: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking—native speech-to-speech models built for real-time voice agents that keep talking while they think, see, and call tools.

This is not cascaded STT → LLM → TTS glue. It’s unfiltered motion from mic to model to speaker, tuned for agents that need to stay conversational under load. Pair it with last month’s Gemini 3.5 Transcribe and the wider Gemini audio suite (Live Translate, Flash TTS, Lyria 3.5), and you’ve got a full generative-audio workbench—not just another chatbot with a wavefile bolted on.

Primary reads: Google’s developer audio post, the 3.8 Live announcement, API docs, and the DeepMind model card.

Native Speech-to-Speech That Stays in the Pocket

Model codes to memorize: gemini-3.8-live (default low-latency dialogue) and gemini-3.8-live-extended-thinking (high-reasoning audio-to-audio). Both speak the Live API language—continuous audio in, audio (and text) out—with huge session room (about 131K input / 65K output tokens) and multimodal context: text, images, audio, and video.

The groove-changers for voice-agent builders:

  • Async function calling — kick tools and APIs in the background while audio keeps streaming, so the agent doesn’t freeze mid-sentence
  • Visual context — ground the chat in what the user sees (cameras, screens, sketches) without breaking conversational swing
  • Alphanumeric precision — confirmation codes, claim numbers, technical strings that used to get mangled in cascaded stacks
  • 97+ languages with accent consistency and mid-conversation language flips
  • Incremental content updates — merge live audio with structured data so replies stay context-aware

Google frames this as a step-change from earlier Live models and a cleaner alternative to cascaded architectures. For producers shipping voice UX, that means fewer brittle handoffs and more “keep the take rolling” energy.

Extended Thinking: Background Brain, Live Narration

When the request needs multi-step reasoning—bookings, sketch-to-code, business-plan jams—3.8 Live Extended Thinking keeps speaking while it reasons. Configure thinking_level as low / medium / high via thinking_config (minimal isn’t supported). Google says it ranks #1 on Artificial Analysis’ Speech-to-Speech leaderboard, with strong agentic scores on τ-Voice-style benches and near-real-time progress narration (“let me check that…”) so users feel the pocket instead of dead air.

Important Live API nuance: on Extended Thinking, turnComplete: true no longer means idle—background reasoning or tool calls may still be cooking. Watch interaction_status (IN_PROGRESS vs IDLE). Function calling is async-only (NON_BLOCKING); blocking mode errors out. Regular gemini-3.8-live skips thinking_level entirely and defaults async tools while still allowing blocking for compatibility.

Pro Tip

Treat Extended Thinking like a parallel send in your DAW: main vocal stays live, background stems render without muting the take. In session setup, start at thinking_level: medium, log when narration feels chatty vs useful, then bump to high only for multi-hop tasks (tool chains, visual+voice workflows). If your client waits on turnComplete alone, you’ll cut the stream early—wire interaction_status: IDLE as the real “ready for next take” gate.

Pricing, Partners, and the Wider Audio Suite

Live API pricing lands around ~$0.005/min audio input and ~$0.018/min audio output (Google’s estimate from token rates)—competitive enough to prototype voice agents without treating every minute like a mastering bill. SynthID watermarking rides on generated audio for transparency.

Integration partners handling real-world media streaming: Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel, Vision Agents. Enterprise and consumer rollouts also touch Gemini Enterprise (private preview), Search Live, Gemini Live, and Workspace Live surfaces (Docs / Gmail / Keep depending on tier).

Suite context (not the hero, but the rack):

  • Gemini 3.5 Transcribe — 85+ languages; WER about 4.0% streaming / 2.6% non-streaming; custom vocabulary, smart polish mode, Interactions API for long files
  • Gemini 3.5 Live Translate — speech-to-speech across 70+ languages
  • Gemini 3.1 Flash TTS — configurable speech generation
  • Lyria 3.5 — production-grade music generation (yes, still in our lane)

How to Get Started

Practical path for voice/audio builders:

  1. Read the stack — Skim the developer audio post and the 3.8 Live announcement so you know Live vs Extended Thinking vs Transcribe.
  2. Try in AI Studio — Open ai.studio/live and run a short speech-to-speech jam before you write infrastructure.
  3. Pick the model string — gemini-3.8-live for fluid agents; gemini-3.8-live-extended-thinking when multi-step background reasoning is the point.
  4. Wire Live API features — enable async tools, decide if you need visual frames (they’re on by default—send only when needed for cost), and plan client logic for interaction_status.
  5. Lean on partners — LiveKit, Pipecat, Vercel, Agora, and friends if you don’t want to own WebRTC plumbing.
  6. Add Transcribe where text is the stem — captions, analytics, speaker-labeled archives via Gemini 3.5 Transcribe / Interactions API.
  7. Keep music in the same ecosystem — when the agent needs beds or motifs, route to Lyria 3.5 instead of bolting on an unrelated music API mid-session.

Gemini 3.8 Live producer cheat sheet (Sep 15, 2026)Models: gemini-3.8-live · gemini-3.8-live-extended-thinkingSuperpowers: async tools · visual context · alphanumeric precision · 97+ langsThinking: low / medium / high (Extended Thinking only)Live API $: ~$0.005/min in · ~$0.018/min outPartners: Agora · Fishjam · LiveKit · LangChain · Pipecat · Vercel · Vision AgentsSuite: 3.5 Transcribe · Live Translate · Flash TTS · Lyria 3.5Try: https://ai.studio/liveDocs: https://ai.google.dev/gemini-api/docs/models/gemini-3.8-live

Conclusion

Gemini 3.8 Live just put native speech-to-speech agents in the center of the mix—async tools that don’t kill the groove, visual grounding, multilingual swing, and Extended Thinking that narrates the deep work while the vocal stays live. Price it by the minute, ship through Live API or a streaming partner, and keep Transcribe / TTS / Lyria on the same rack when you need text stems or music beds.

Open ai.studio/live, push a messy real-world prompt, and listen for whether the agent stays in the pocket when the tools start firing. That’s the heavy impact test.

—Anabel ♡