Перейти к основному содержимому

Gemini 3.8 Live and Live Extended Thinking: The Builder's Field Guide to Google's New Voice Models (Sep 2026)

Средний

On September 15, 2026 Google released two speech-to-speech models: Gemini 3.8 Live (gemini-3.8-live) and Gemini 3.8 Live Extended Thinking (gemini-3.8-live-extended-thinking). Both are generally available today in the Gemini API and Google AI Studio. Google's launch post is mostly benchmarks and demos. If you are building a voice agent, the details that matter are in the developer docs: tools now run asynchronously by default, one config field now returns an error, and one feature is always on and billed while the model listens. This page collects those details in one place.

What you'll learn
  • Pick between the two models using the real API differences, not the marketing names
  • Migrate a gemini-3.1-flash-live-preview agent without hitting the thinkingLevel, affective-dialog, and blocking-tool traps
  • Design around the session walls: 15 min audio-only, 2 min audio+video, ~10 min per WebSocket connection
  • Estimate what a voice session really costs, and why cost grows faster than minutes
  • Ship a browser client without leaking an API key

The one-sentence version​

Use gemini-3.8-live as the default for low-latency voice agents. Use gemini-3.8-live-extended-thinking when the caller asks hard, multi-step questions and a short delay is acceptable. Both cost $3.00 / 1M audio input tokens (~$0.005/min) and $12.00 / 1M audio output tokens (~$0.018/min), plus $0.75 / $4.50 for text. Both have a 131,072-token input window and 65,536-token output limit, and both have a free tier in the Gemini API.

What shipped, and what Google claims​

gemini-3.8-livegemini-3.8-live-extended-thinking
Google's positioning"Default option for most low-latency voice agent experiences""High-reasoning audio-to-audio model supporting background reasoning during live audio interactions"
ThinkingInterleaved reasoning; thinkingLevel is not acceptedthinkingLevel = low / medium / high (no minimal); includeThoughts: true for summaries
Function callingAsync by default (NON_BLOCKING); BLOCKING still available; scheduling INTERRUPT / WHEN_IDLE / SILENTAsync only, with no blocking mode and no scheduling control
Consumer surfacesSearch LiveGemini Live app; Docs, Gmail, Keep for subscribers
Headline numbers (Google)2nd on Artificial Analysis' Speech Agent Arena82.6 Speech-to-Speech Quality Index (1st); 68.6% τ-Voice; 35.1% τ-Voice-banking; 97.7% Big Bench Audio

Look at the 35.1% on τ-Voice-banking, the number Google chose to publish. Even the top-ranked voice model completes only about a third of realistic banking tasks by voice. Treat that as the ceiling for unattended voice agents that handle money, and put confirmation gates on anything consequential (the same rule as in Full-duplex voice AI).

Every audio output is watermarked with SynthID. If your product records or re-publishes model speech, AI-audio detectors can identify it.

Three things most coverage misses​

1. The "thinking" model narrates while it reasons​

Extended Thinking does not go silent while it thinks. Google says it "reasons and speaks simultaneously." It gives early verbal cues ("let me check that") and narrates progress while reasoning continues in the background. That changes how your client decides when the model is finished. On this model turnComplete: true does not mean the session is idle. Background reasoning can still be running. Check interaction_status instead (IN_PROGRESS vs IDLE) before you end a call, hand off to a human, or start the next scripted step.

This is the same "fast talker, slow thinker" split OpenAI introduced with GPT-Live. The difference is that Google puts both roles in one model, where GPT-Live delegates the reasoning to a separate frontier model.

2. Upgrading from 3.1 Flash Live changes behavior, not just the model string​

Google's migration notes for gemini-3.1-flash-live-preview → gemini-3.8-live include four changes that alter what your code does:

  • Remove thinkingLevel. 3.1 Flash Live defaulted to minimal. 3.8 Live does not accept the field, so a config copied from 3.1 must drop it.
  • Tool calls are now asynchronous by default. On 3.1 Flash Live, "the model will not start responding until you've sent the tool response." On 3.8 Live the model keeps talking while your function runs. That is good for latency, but a tool that must finish first (an identity check, a payment) now needs behavior: BLOCKING set explicitly. The Extended Thinking model has no blocking mode, so those flows need a gate in your own code.
  • Proactive audio is always on. The model can decide not to respond when speech isn't addressed to it. You cannot turn this off, and the best-practices guide notes that you are billed input tokens while it listens.
  • Affective dialog is gone. If your agent relied on enableAffectiveDialog to adapt to the caller's tone, remove it and test the new default behavior.

Two changes help you. send_client_content now works for the whole session, so you can inject text context (CRM data, a transcript of a transfer) mid-call. Turn coverage now defaults to TURN_INCLUDES_AUDIO_ACTIVITY_AND_ALL_VIDEO.

3. A voice session's cost grows faster than its length​

Audio accumulates at roughly 25 tokens per second (about 1,500 per minute). That is where the per-minute prices come from: 1,500 × $3/1M ≈ $0.0045 in, and 1,500 × $12/1M = $0.018 out. The trap is that each turn re-processes the entire accumulated context, not only the new audio. The best-practices guide says so directly. Two smaller costs stack on top: enabling transcription bills text output on top of audio output, and proactive audio bills input tokens while the model listens.

A rough illustration (our arithmetic, not a Google figure): take a 10-minute call with 20 turns. The context grows from 0 to about 15,000 tokens, so the average turn re-reads about 7,500 tokens of input. That is about 150,000 input tokens (~$0.45), compared with ~$0.045 if only new audio were counted. Longer calls cost more per minute, not just more in total. Context window compression, which you need anyway to get past the 15-minute wall, also limits this cost growth.

The session walls, and how to design around them​

Guided walkthrough1 of 6
  1. Without compression, audio-only sessions stop at 15 minutes and audio+video sessions at 2 minutes. Separately, each WebSocket connection lasts around 10 minutes. The 2-minute video limit catches most screen-share demos.

Browser and mobile clients: never ship the API key​

Live sessions work best when the client connects directly to Google, because proxying real-time audio through your backend adds latency. The safe way to do this is an ephemeral token. Your server mints it, and the client uses it once. Defaults: the token expires 30 minutes after creation, it must be used to start a new session within 1 minute, and it allows 1 use. You can lock it with liveConnectConstraints to a specific model, response modality and resumption settings, so a stolen token cannot be used for anything else. Ephemeral tokens work only with the Live API and only on v1beta.

System-instruction skeleton for gemini-3.8-live (async tools, always-on proactive audio)

You are Ava, a voice assistant for Northwind Travel. You are SPEAKING, not writing.
RESPOND IN ENGLISH. YOU MUST RESPOND UNMISTAKABLY IN ENGLISH.

Style:
- Short sentences, one idea each. Never read lists, markdown, or URLs aloud.
- If speech is clearly not addressed to you (TV, side conversation), stay silent.

Tools:
- lookup_booking: invoke ONLY after the caller has given BOTH booking reference AND last name.
- While a lookup is running, tell the caller you are checking. Do not guess the result.
- change_booking: before invoking, read back the exact change and wait for a clear spoken "yes".

Safety:
- Repeat back every number, date, and name before acting on it.
- If audio is unclear, say what you think you heard and ask.
- Anything involving refunds or payment details: offer a human agent instead.

The skeleton follows Google's own guidance for Live agents: state the output language explicitly in capitals, and give each tool a precise condition for when to call it. That second rule matters more with async tools, because the model talks while the tool runs and will fill the silence with a guess unless you tell it not to.

Gemini 3.8 Live vs the alternatives​

  • vs Claude. Claude's voice features remain turn-based: voice mode in the Claude apps, push-to-talk in Claude Code. A common hybrid uses a Live model as the conversational front end and calls Claude as a tool for heavy reasoning or coding. Async function calling suits this pattern well, since the call can keep going while the tool runs.
  • vs OpenAI. GPT-Live is a speech-native, full-duplex ChatGPT product. For developers, OpenAI's offering is gpt-realtime on the Realtime API (see Full-duplex voice AI for the landscape).
  • Early community reports (Hacker News launch thread, unverified). Commenters praised latency, voice quality, accent handling and live translation. Others said Extended Thinking reasoned inconsistently and that the model "invents additional requirements" in agentic flows. These are anecdotes, not measurements, but they match the 35.1% banking score: test your own task end to end before trusting it.

Check yourself

0/5
  1. You port a gemini-3.1-flash-live-preview config that sets thinkingLevel: minimal to gemini-3.8-live. What should you do?
  2. A payment-verification tool must finish before the agent says anything else. Which setup is correct?
  3. On gemini-3.8-live-extended-thinking you receive turnComplete: true. Can you safely hang up?
  4. Why does a 20-minute voice call cost more than twice a 10-minute call?
  5. Your screen-sharing support demo keeps disconnecting at about two minutes. Most likely cause?
Gemini Live API vocabulary
Нажмите Enter или пробел, чтобы перевернуть карточку. Используйте стрелки влево и вправо для перехода между карточками.Показан термин.
1 / 8
Key takeaways
  • Two GA models, one price: gemini-3.8-live is the low-latency default; gemini-3.8-live-extended-thinking reasons in the background while it keeps speaking.
  • Migrating from 3.1 Flash Live changes behavior: remove thinkingLevel, mark must-finish tools BLOCKING, expect always-on proactive audio (billed while listening), and drop affective dialog.
  • Plan for the limits up front: 15 min audio-only, 2 min audio+video, ~10 min per connection. Compression, resumption handles and GoAway handling are required for production.
  • Cost grows faster than call length because every turn re-reads the whole context. Compression is both the session-length fix and the cost fix.
  • Even Google's best score on voice banking tasks is 35.1%. Keep spoken confirmation gates on anything that moves money or data.

Sources & further reading​

Next​