Google DeepMind · 2026-09-15 · major
Gemini 3.8 Live — Google's voice models talk while they think
Gemini 3.8 Live is Google DeepMind's new audio-to-audio model for the Gemini Live API. A second variant, Extended Thinking, reasons in the background while it keeps talking, and scores 82.6 on the Speech to Speech Quality Index.
Google's new Live API models hold a real-time voice conversation and keep reasoning in the background while they speak.
Quick facts
| Maker | Google DeepMind |
|---|---|
| Model IDs | gemini-3.8-live, gemini-3.8-live-extended-thinking |
| Availability | Gemini API, Google AI Studio, Search Live, Google Workspace |
| Languages | 97, detected and switched mid-conversation |
| Watermark | SynthID on all generated audio |
| Access | Proprietary, API only; Gemini Enterprise in private preview |
Benchmarks
Pricing
| Audio input · or $0.005 / min | $3.00 / 1M tokens |
|---|---|
| Audio output · or $0.018 / min | $12.00 / 1M tokens |
| Text input | $0.75 / 1M tokens |
| Text output | $4.50 / 1M tokens |
| Image / video input · or $0.002 / min | $1.00 / 1M tokens |
| Free tier · Free of charge for input and output | $0 |
What is it?
Gemini 3.8 Live replaces gemini-3.1-flash-live-preview as Google's audio-to-audio model for the Gemini Live API. It ships in two flavours: the base model for cheap, fast voice agents, and an Extended Thinking variant for harder tasks. Both take streamed audio, video, images and text in, and answer with speech.
How does it work?
Instead of waiting for a turn to end, the Extended Thinking variant runs its reasoning in parallel with speech, filling the gap with natural spoken acknowledgements such as "Let me check that…". Tool calls and API requests run in the background so the audio stream never stalls, visual input is processed in near real time, and language detection happens mid-conversation across 97 languages.
Why does it matter?
Voice agents usually break the illusion in the pause between a question and an answer, and that pause gets worse the harder the question is. By reasoning and talking at once, Gemini 3.8 Live Extended Thinking pushes the top score on the Artificial Analysis Speech to Speech Quality Index to 82.6 and reaches 97.7% on Big Bench Audio, so support and booking flows can use a stronger model without sounding slower.
Who is it for?
developers building voice agents and phone support flows
Frequently asked questions
- What is the difference between Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking?
- Gemini 3.8 Live is built for scale and cost efficiency and is Google's default pick for low-latency voice agents that should not pause to reason. Extended Thinking is the higher-intelligence option for complex, multi-step work; it reasons in the background and speaks at the same time, so the extra thinking does not stop the conversation.
- How much does Gemini 3.8 Live cost?
- Gemini 3.8 Live is metered two ways on Google's Gemini API pricing page: per million tokens or per minute of media. Audio is the main cost at $3.00 in and $12.00 out per million tokens, which Google also quotes as $0.005 and $0.018 per minute. Text and image or video input are cheaper, and the free tier costs nothing.
- Can Gemini 3.8 Live switch languages during a call?
- Yes. Gemini 3.8 Live covers 97 languages and detects which one the speaker is using, then switches mid-conversation rather than being locked to the language chosen when the session opened. A caller can start in one language and move to another without restarting, which matters for support lines that serve mixed-language regions.
- Is audio from Gemini 3.8 Live marked as AI-generated?
- Google applies SynthID watermarking to all audio that Gemini 3.8 Live generates. SynthID embeds a signal into the waveform that survives normal playback and re-encoding, so a recording of a Gemini voice agent can later be checked as machine-generated. The watermark is applied by default rather than being a setting developers turn on.
Try it
Call model id `gemini-3.8-live` on the Gemini Live API, or try it in Google AI Studio.