Supervised customer support
Voice experiences that authenticate users, answer bounded questions, call narrow read-only tools, and quickly hand off to a human for exceptions or consequential requests.
Independent tool overview
Gemini 3.1 Flash Live at a glance
Gemini 3.1 Flash Live Preview is Google's low-latency audio-to-audio model for real-time voice and multimodal agents. It can listen to audio, receive text, images, and video, respond with speech or text, use Google Search grounding, and call synchronous functions. It is compelling for conversational support, coaching, and guided workflows, but it remains a preview with important session, feature, privacy, consent, and cost constraints.
Visit the official Gemini 3.1 Flash Live site ↗
- Model ID
- gemini-3.1-flash-live-preview
- Lifecycle
- Preview
- Inputs
- Text, images, audio, and video
- Outputs
- Text and audio
- Token limits
- 131,072 input and 65,536 output
- Agent tools
- Synchronous function calling and Google Search grounding
- Reviewed
- August 31, 2026
Overview
What Gemini 3.1 Flash Live is
Gemini 3.1 Flash Live powers real-time dialogue in Google's Gemini Live and Search Live experiences and is available to developers through the Gemini Live API under the model ID `gemini-3.1-flash-live-preview`. Google also offers it in Gemini Enterprise for Customer Experience.
The model is designed to understand acoustic cues such as pitch, pace, interruptions, hesitation, and noisy input while maintaining a natural spoken response. It accepts streaming text, images, audio, and video, produces text or 24 kHz audio, supports thinking levels, Google Search grounding, and synchronous function calling.
This is not a general Gemini endpoint with every tool. The current model page says code execution, file search, URL context, structured output, image generation, caching, Maps grounding, and batch processing are not supported. Asynchronous function calling, proactive audio, and affective dialogue are also absent in this model.
A production voice agent is a safety and operations system, not just a prompt. It needs disclosure, consent, authentication, escalation, tool allowlists, spend controls, interruption handling, transcript and recording rules, adversarial-audio testing, and a clear response when the model is uncertain or the connection fails.
Use cases
Who Gemini 3.1 Flash Live is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Interactive coaching and assistance
Language practice, product guidance, accessibility, brainstorming, and step-by-step help where low latency and tone awareness matter but a mistaken answer is recoverable.
Multimodal field workflows
Applications that combine a live conversation with camera or screen context for troubleshooting, inspection, or guided tasks under human control.
Voice-agent prototyping
Teams evaluating turn-taking, interruption behavior, multilingual quality, tool calling, latency, cost, and user acceptance before committing to a production architecture.
Capabilities
Core Gemini 3.1 Flash Live features
Native audio-to-audio conversation
Streams speech directly to and from the model for more natural timing and vocal nuance than a simple speech-to-text, text-model, and text-to-speech chain.
Multimodal awareness
Accepts text, still images, streaming audio, and video so an agent can discuss what a user says and shows in the same session.
Turn detection and interruption
Supports automatic or custom voice-activity detection and streaming turn control for conversational back-and-forth. Applications must tune sensitivity and silence thresholds for their environment.
Thinking levels
Uses `minimal`, `low`, `medium`, or `high` thinking levels, defaulting to minimal for low latency. More reasoning can improve hard tasks but increases time and token cost.
Synchronous function calling
Can request application tools during a conversation, but waits for the tool result before responding and does not support the model's asynchronous nonblocking tool mode.
Google Search grounding
Can query Google Search for current information. Each user request can trigger multiple billable search queries, and grounding data has separate retention terms.
SynthID audio watermarking
Google says all audio produced by 3.1 Flash Live carries an imperceptible SynthID watermark intended to support detection of AI-generated audio.
Session resumption and compression
Can resume across periodic WebSocket reconnections and compress the context window to support longer sessions, with explicit cost and data-retention tradeoffs.
Process
How the Gemini 3.1 Flash Live workflow works
Step 1
Define the allowed conversation
State what the agent may discuss, what tools it can use, what it must never decide or change, how it discloses AI use and recording, and the exact triggers for human or emergency escalation.
Step 2
Secure the connection
Authenticate the user on your backend and issue a constrained short-lived ephemeral token for direct client WebSocket access. Never ship a long-lived Gemini API key in a browser or mobile app.
Step 3
Minimize and label data
Collect only necessary audio, video, images, transcripts, and identifiers; obtain required consent; document recording and retention; redact sensitive fields; and keep regulated data out unless the governing contract and controls explicitly allow it.
Step 4
Constrain tools and identity
Use server-side authorization, narrow schemas, read-only defaults, amount and frequency limits, confirmation for external effects, and fresh identity checks before exposing accounts, placing orders, making payments, scheduling, or modifying records.
Step 5
Test real conversations
Evaluate accents, languages, background noise, cross-talk, silence, interruptions, children, distress, numbers, addresses, names, prompt injection spoken aloud or shown on camera, tool failures, network drops, and repeated reconnects.
Step 6
Control context and cost
Set context-window compression, monitor input and output audio minutes, account for transcription and grounding surcharges, handle GoAway messages, resume connections, and stop idle sessions promptly.
Step 7
Review and improve safely
Measure task completion, escalation accuracy, latency, hallucinations, unsafe tool attempts, demographic quality gaps, cost, abandonment, complaints, and consent failures without retaining more raw conversation data than necessary.
Cost
Gemini 3.1 Flash Live pricing and free plan
The paid Gemini Developer API charges separately for text, image or video, audio, and output. For a rough audio-only minute, published rates are $0.005 per minute of input and $0.018 per minute of generated audio before context reprocessing, transcripts, thinking, Search grounding, and application infrastructure. Persistent context compounds billing across turns, so a long conversation costs more than simply multiplying wall-clock minutes.
Gemini API free tier
No token charge within available quota
For testing in AI Studio and limited developer evaluation, subject to eligibility and changing preview quotas.
- Google's pricing table marks free-tier data as used to improve products
- Current usable audio quota may require an active billing project even when the pricing table shows a free tier
- Preview limits can be restrictive or change
- Do not use confidential production conversations under consumer-style evaluation terms
Paid text input
$0.75 per 1M tokens
Text instructions, conversation context, and other text sent to the live model.
- Context can be billed again on later turns
- System instructions and tool content contribute to usage
- Thinking and generated transcript text are billed on the output side
- Paid-service prompts are not used to improve Google's products
Paid audio input
$3 per 1M tokens or about $0.005/minute
Streaming microphone or other audio sent while input is active.
- The model is billed only while actively streaming input because proactive audio is unsupported
- Audio accumulates at roughly 25 tokens per second
- Long retained context can increase later-turn cost
- Silence and always-on capture should still be minimized
Paid image and video input
$1 per 1M tokens or about $0.002/minute
Visual context supplied during the live session.
- The default turn can include all video frames while audio is active
- Limit frame rate and send visual context only when useful
- Audio-video sessions hit the uncompressed context limit much sooner
- Images and video raise privacy and bystander-consent concerns
Paid text output
$4.50 per 1M tokens
Text responses, thinking tokens, and enabled transcription text.
- Input and output audio transcription is a surcharge at the text output rate
- Thinking level affects generated token use
- Do not enable transcripts merely for convenience if they are not needed
- Transcripts create a separate sensitive data asset
Paid audio output
$12 per 1M tokens or about $0.018/minute
Model-generated spoken responses.
- All 3.1 Flash Live audio is SynthID-watermarked according to Google
- Verbose responses directly increase latency and cost
- Context reprocessing and other modalities are additional
- Audio playback, telephony, and transport costs are outside this model price
Google Search grounding
5,000 shared searches/month, then $14 per 1,000
Optional current-web grounding shared across Gemini 3.x models.
- A single customer request can create multiple search queries
- Google stores prompts, context, and output for 30 days when Search grounding is used
- Search evidence can still be wrong or misapplied
- Disable grounding for conversations that require a strict zero-data footprint
Pricing checked . Check current pricing at the source ↗
Assessment
Gemini 3.1 Flash Live strengths and limitations
Where it stands out
- Native low-latency speech preserves conversational rhythm and acoustic context better than a loosely coupled transcription pipeline.
- Text, image, audio, and video input enable genuinely multimodal guidance instead of voice-only question answering.
- Configurable thinking lets developers favor quick turn-taking or allocate more reasoning to a complex request.
- Function calling turns conversation into bounded application workflows when tools are carefully authorized.
- Search grounding can supply current context without leaving the live conversation.
- Ephemeral tokens enable lower-latency direct client connections without embedding a permanent API key.
- Google publishes detailed session, migration, pricing, and zero-data-retention guidance that makes the major tradeoffs visible.
What to consider
- Both the model and Live API are in preview, so behavior, quotas, supported capabilities, prices, and lifecycle can change before general availability.
- Without context compression, audio-only sessions are limited to about 15 minutes and audio-video sessions to about two minutes; individual WebSocket connections last around ten minutes and require resumption for continuity.
- Billing compounds because retained conversation context is processed on later turns. Transcription, thinking, grounding, telephony, storage, and infrastructure can make real cost materially higher than the per-minute audio headline.
- Asynchronous function calling is unavailable. The model waits for each tool response, which can create awkward silence and latency during slower operations.
- Code execution, file search, URL context, structured output, caching, image generation, Maps grounding, batch use, proactive audio, and affective dialogue are not supported by this endpoint.
- Voice agents can mishear names, numbers, negation, consent, account details, or urgent context. Never use an unconfirmed transcript to make a consequential change.
- Spoken, visual, or retrieved prompt injection can manipulate tool use. Server-side authorization and confirmation must control actions regardless of what the model requests.
- Audio, video, camera frames, transcripts, search queries, tool results, and session state can contain highly sensitive or biometric-adjacent information. Recording and privacy laws vary by jurisdiction and participant.
- Paid API data is not used to improve Google's products, but paid prompts and responses can still be logged temporarily for abuse monitoring. Session resumption retains text, audio, and video state for up to 24 hours, and Search grounding stores related data for 30 days.
- The model must not impersonate a person, obscure that it is AI, or provide autonomous medical, legal, financial, emergency, hiring, education, or eligibility decisions. SynthID is a useful signal, not a complete disclosure or misuse control.
Compare
Gemini 3.1 Flash Live alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Miscellaneous
GPT-Realtime-2
Choose GPT-Realtime-2 to evaluate OpenAI's current real-time voice model, tool behavior, safety controls, pricing, and ecosystem against Google's Live API.
Explore GPT-Realtime-2 →Agents
Grok Voice Agent
Choose Grok Voice Agent when xAI's voice model, tool calling, and real-time data stack better fit the application and governance requirements.
Explore Grok Voice Agent →Sales
Vapi
Choose Vapi when the bigger need is provider-agnostic voice-agent orchestration, telephony, routing, observability, and swapping underlying speech or language models.
Explore Vapi →Marketing
ElevenLabs
Choose ElevenLabs when voice quality, voice design, speech generation, and its broader audio platform matter more than using Google's native multimodal model.
Explore ElevenLabs →Questions
Gemini 3.1 Flash Live FAQs
What is Gemini 3.1 Flash Live?
It is Google's preview low-latency audio-to-audio model for real-time voice and multimodal applications. It accepts text, images, audio, and video; returns text or speech; and supports thinking, Google Search grounding, and synchronous function calling.
How much does Gemini 3.1 Flash Live cost?
Published paid rates are $0.005 per minute of audio input and $0.018 per minute of audio output, or token-based equivalents, plus any text, visual input, thinking, transcription, context, grounding, and infrastructure charges.
Does Gemini 3.1 Flash Live have a free tier?
The pricing table shows free usage within available quota and says free-tier data can improve Google products. Preview availability can require an active paid project or change, so verify the actual project quota before designing around free access.
Can Gemini 3.1 Flash Live call tools?
Yes, through synchronous function calling. It waits for the tool result before continuing. The application must independently authenticate, authorize, validate, rate-limit, and confirm any consequential action.
How long can a Live API conversation last?
Without compression, about 15 minutes for audio-only and two minutes for audio-video. Context compression can extend a session, while session resumption handles WebSocket connections that reset around ten minutes. Both features affect cost or retention.
Can I put the Gemini API key in a web or mobile app?
No permanent key should be shipped client-side. Google recommends authenticating the user with your backend and issuing a short-lived, constrained ephemeral token for a direct client-to-Live-API WebSocket.
Does Google train on Gemini Live API conversations?
The pricing table says free-tier data is used to improve products and paid-tier data is not. Paid traffic can still be logged temporarily for abuse monitoring. Data retention also changes if Search grounding or session resumption is enabled.
Is Gemini 3.1 Flash Live zero data retention?
Not automatically. Google says a paid project needs approved ZDR controls, must avoid session resumption to prevent up-to-24-hour state retention, and cannot use Search grounding if it requires zero retention because grounding stores data for 30 days.
Is it safe for customer service or healthcare calls?
Only with an appropriate contract, jurisdictional consent, strict data minimization, identity checks, bounded tools, human escalation, quality testing, and qualified legal and security review. A capable voice model alone does not establish HIPAA, financial, recording-law, or consumer-protection compliance.
Bottom line
Our Gemini 3.1 Flash Live verdict
Gemini 3.1 Flash Live is a strong foundation for natural, multimodal voice experiences when a team wants Google's models, grounding, and function-calling stack. The product decision should hinge less on a demo voice and more on production controls: preview tolerance, session engineering, compounding context cost, ephemeral authentication, consent, data retention, tool authorization, and human escalation. It is well suited to bounded assistance; it should not autonomously make consequential decisions or actions.
Visit Gemini 3.1 Flash Live website ↗Similar Business Operations Tools
View all Business Operations AI tools →
Voxtral TTS
Voxtral TTS - Mistral's voice cloning model for building multilingual AI speech agents
Lovable Payments
Lovable Payments - Lovable's one-conversation setup for adding payments, tax, and multi-currency support to AI-built apps
Claude Marketplace
Claude Marketplace - Anthropic's enterprise hub for buying Claude-powered partner tools
Cofounder 2
Cofounder 2 - General Intelligence's agent orchestrator that manages full business operations across engineering, sales, marketing, and design
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.