Gemini 3.8 & 3.8 Live Extended Thinking
Google DeepMind · Voice & Language · Chatbot
Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are two real-time voice AI models from Google DeepMind, both launched on September 15, 2026. The standard 3.8 Live is a speech-to-speech model tuned for fast, low-cost voice interfaces, while 3.8 Live Extended Thinking adds background multi-step reasoning so it can keep talking while it works through a hard problem. Developers reach both through the Gemini API, Google AI Studio, and the Gemini Live API. No extra setup. Regular users are already meeting them inside Gemini Live and Search Live, and Google says both models have started rolling out across the Gemini app and Google Workspace too.
About Gemini 3.8 & 3.8 Live Extended Thinking
What Is Gemini 3.8 & 3.8 Live Extended Thinking
These are Google's most advanced real-time conversation models to date, according to the company. Both are natively speech-to-speech, which means they hear audio and answer with audio instead of running your words through a separate transcription model, then a text model, then a voice generator. That older pipeline is where a lot of delay and tone loss came from.
The split between the two comes down to how they handle a tough question, and that one design choice shapes which one you should build on. Gemini 3.8 Live answers fast and calls tools in the background, which suits customer support routing, voice search, and language practice. Gemini 3.8 Live Extended Thinking keeps a "thinking" process running while it speaks, so it can narrate progress on multi-step tasks like trip planning or debugging code without going silent.
The main limits are practical. Extended Thinking only supports non-blocking asynchronous tool calls. You also have to wait for an explicit idle signal before ending a session. And the highest thinking levels cost more per minute of audio. Pricing runs on audio minutes, not tokens, so heavy always-on voice apps need budget math up front.
Getting Started
- Open Google AI Studio or the Gemini API console and pick either
gemini-3.8-liveorgemini-3.8-live-extended-thinking. - For live voice apps, connect through the Gemini Live API over a stateful WebSocket rather than a single request.
- Send input audio as 16-bit PCM at 16 kHz little-endian, and expect output audio at 16-bit PCM, 24 kHz.
- If you use Extended Thinking, set a thinking level (low, medium, or high) and use only non-blocking tool calls.
- End the session only after you see an idle status from the model.
Product Information
A quick look at Gemini 3.8 & 3.8 Live Extended Thinking's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Developers building voice agents
- Customer support teams
- Regular Gemini app users
Tasks
- Real-time troubleshooting
- Live language practice
- Booking and scheduling
Scenarios
- Running a hands-free assistant while driving or cooking, where waiting in silence for an answer feels broken.
- Coaching someone through a whiteboard sketch, since 3.8 Live reads visual input near real time.
- Coordinating several async functions in one spoken session, such as checking inventory, then price, then delivery time.
Key features
Speech-to-speech conversation
Both models skip the old transcribe-then-generate chain. Audio goes in. Audio comes back. The payoff is fewer dropped cues in tone and less lag between what you say and what comes back. For a voice app, that gap is the whole experience.
Thinking out loud without going silent
Extended Thinking runs its reasoning in parallel with speech. Instead of dead air on a hard question, it opens with a short verbal hold like "Let me check that," then narrates progress as it works. Google's demos include live chess guidance and turning whiteboard sketches plus spoken feedback into a working React component.
Background tool and API calls
You can keep talking while the model runs a web search or calls a tool. In Extended Thinking, only non-blocking asynchronous tools are supported. That last detail matters. A blocking call would stall the very flow the model is trying to protect.
97-language switching in one call
Gemini 3.8 Live detects and switches among 97 supported languages without you restarting the session. Multilingual households and support desks that field mixed-language calls get the most out of this, and nobody has to pick a language up front.
Visual input in near real time
The standard model understands images and video frames as they come in, roughly one frame per second. That lets it comment on what a camera sees.
SynthID audio watermarking
Every audio clip these models generate carries an invisible SynthID watermark. There's no toggle to disable it. It won't change what you hear, but it does give platforms a way to flag AI-generated speech.
Pros and cons
Pros
- Speech-to-speech design cuts the delay that made older voice assistants feel robotic.
- Extended Thinking keeps a conversation alive while it works, so complex requests don't trigger silent dead air.
- Per-minute audio pricing of $0.005 in and $0.018 out stays predictable for high-volume apps.
- Native switching across 97 languages suits multilingual products without extra setup.
- Invisible SynthID watermarking comes built in for anyone shipping generated audio publicly.
Cons
- Extended Thinking only allows non-blocking async tool calls, so synchronous tools force a different architecture.
- You must wait for an explicit idle status to end a session, which adds a state your code has to track.
- Audio-minute billing can climb fast for always-on assistants, since there's no flat cap.
- The strongest thinking levels raise per-minute cost, so the deepest reasoning isn't free.
Frequently asked questions
Gemini 3.8 Live is tuned for fast, low-cost real-time conversation, while 3.8 Live Extended Thinking adds background multi-step reasoning and can speak and think at the same time. They share the same speech-to-speech foundation, so pick based on latency versus depth.
Related content
Explore related tools, skills, and articles for Gemini 3.8 & 3.8 Live Extended Thinking.
Gemini 3.8 & 3.8 Live Extended Thinking Alternatives
Foley
Hanamori Labs, LLC · Voice & Language · Productivity · AI AssistantFoley is a voice to-do list that turns a spoken sentence into a filed task, complete with project, date and time, then helps you plan the day around what is already there.
HireKey
HireKey · Voice & Language · Productivity · LeaningHireKey is an AI job search copilot built by a former tech recruiter. It checks role fit, tailors resumes, runs interview practice and tracks every application.
Fluence
Yay Software (Yannick Huchard) · Writing · Voice & Language · ProductivityFluence is an Android app that reads PDFs, EPUBs, Word files and web articles aloud, follows each sentence as it speaks, and answers questions about the text on the phone itself.