The Gemini Live API streams your voice (and optionally camera) to Google and streams spoken answers back, with interruptions handled for you. The current model is Gemini 3.8 Live, released 15 September 2026. Audio costs about half a cent per minute in and two cents out. It's a developer tool; expect a day of setup, not an hour.
Other good fits
Voice apps that also need to look up facts or call your own functions while talking, such as checking a booking or reading a stock level.
+ Pros
Read pros
The Live API is the part of Google's Gemini API for spoken, real-time conversation. Your app keeps a connection open, streams microphone audio (and optionally video) to Google, and receives spoken audio back while the user is still in the conversation. It detects when someone starts talking and stops speaking when interrupted.
This catalogue entry is about the developer API. The voice mode in the Gemini phone app is also called Gemini Live, but it's a feature of the Gemini app, not something you build with.
The current models, both released on 15 September 2026, are Gemini 3.8 Live (gemini-3.8-live, the default for low-latency voice) and Gemini 3.8 Live Extended Thinking (for when the assistant must reason before answering). The older gemini-3.1-flash-live-preview is now labelled legacy.
Pick it when:
You're building a voice front end: an assistant that answers questions about opening hours, sample drop-off rules or a device manual, out loud.
The assistant must do something while talking: call your function to look up an order, or search the web, then answer.
You want speech in and speech out from one model, instead of chaining separate speech-to-text, chat and text-to-speech services.
− Cons
Read cons
It's plumbing, not a product. Audio capture, echo cancellation, reconnects and a phone line are yours to build.
Session limits need handling from day one, or conversations cut off.
Voice mistakes are harder to catch. Nobody re-reads a spoken answer. Keep the assistant's scope narrow and tell it to admit when it doesn't know.
Personal data in speech goes to Google's cloud. Tell callers they're talking to an AI, and check your data protection rules before recording anything.
Your team already builds on OpenAI: OpenAI Realtime does the same job.
You want to swap models and speech providers freely: build the pipeline with Pipecat, which also supports the Gemini Live API.
Avoid if
You want a phone agent without writing code. A hosted voice-agent platform such as ElevenLabs or Retell is quicker.
Calls contain health, HR or other sensitive personal data and you have no data processing agreement covering Google's cloud.
Context
Live behavior depends on the selected current model, client transport and session configuration.
Requirements
Use a supported Google API project, credential and Live API client flow. Microphone permissions, audio capture, network recovery and any client-side storage remain application responsibilities.
Licence & cost
Read licence & cost
Prices checked 23 September 2026 (Gemini API, paid tier, Gemini 3.8 Live and 3.8 Live Extended Thinking):
Text
Audio
Image/video
Input
$0.75 per 1M tokens
$3.00 per 1M tokens (about $0.005 per minute)
$1.00 per 1M tokens (about $0.002 per minute)
Output
$4.50 per 1M tokens
$12.00 per 1M tokens (about $0.018 per minute)
–
At the per-minute rates, ten minutes of listening plus five minutes of the model speaking comes to about 14 cents, before tool calls. There's a free tier for testing. Outside Europe Google may use free-tier content to improve its products; in the EEA, Switzerland and the UK paid-tier data rules apply to all use, and apps offered to users there must run on the paid tier.
The running effort is bigger than the price. Sessions end after 15 minutes of audio (2 minutes with video) unless you switch on context compression, and each connection lasts about 10 minutes, so your app has to resume sessions. In a browser or phone app, use short-lived ephemeral tokens, never your real API key.
First task
Documentation-based suggestion — not run here.
Read first task
Build a spoken FAQ for one topic and test it with five questions. Allow two to three hours; you need Python, a microphone, headphones (to stop the model hearing itself) and a Gemini API key from Google AI Studio (opens in a new tab).
Try it first without code: AI Studio has a Live mode where you can talk to the model in the browser.
Copy the Python microphone example from the Live API guide. Set the model to gemini-3.8-live.
Replace the system instruction with a short FAQ, for example ten rules for handing in samples at a lab (drop-off times, labelling, what isn't accepted). Add: "Only answer from these rules. If the answer isn't there, say so."
Ask five questions out loud: three the rules answer, one they don't, and one where you interrupt the answer halfway.
Success check: the three answerable questions get correct spoken answers, the unanswerable one gets "I don't know" rather than an invented rule, and the model stops talking when you interrupt. For the bigger picture of voice pipelines, see T09-L03.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Checked against Google's Live API guide, session-management guide, Gemini API model list and 3.8 Live model page, pricing page, deprecation schedule and API terms on 23 September 2026. We did not open a Live session for this review.
Live API guide · Sep 23, 2026 — Real-time voice and video over WebSockets; voice activity detection, tool use, session management and ephemeral tokens; 16 kHz PCM input, 24 kHz output.
Gemini API pricing · Sep 23, 2026 — 3.8 Live: input $0.75 text, $3.00 per 1M tokens or $0.005/min audio; output $4.50 text, $12.00 or $0.018/min audio.
Gemini deprecations · Sep 23, 2026 — gemini-3.8-live and gemini-3.8-live-extended-thinking released 15 September 2026; 3.1 Flash Live is a legacy preview.
Live API session management · Sep 23, 2026 — Audio-only sessions limited to 15 minutes, audio-video to 2 minutes without context compression; connections last about 10 minutes.
Gemini 3.8 Live model page · Sep 23, 2026 — 131,072 input tokens; text, image, audio and video in; text and audio out; function calling and Search grounding supported.
Further review notes & alternatives
Alternatives and trade-offs
OpenAI Realtime wins when you already have OpenAI contracts and keys. Same idea: one streaming connection, speech in and out.
ElevenLabs' voice agents win for a finished voice agent with a phone number and many voices, configured in a web dashboard.
Pipecat wins when you want control: an open-source framework that can use Gemini Live or other speech and model services.
Stream audio in and receive spoken responses for voice applications.
Combine conversation with tools, grounding, and multimodal inputs.
Use unified token billing when a Gemini API workflow fits.