Gemini 3.8 Live: Voice Agents Get Real (and Cheaper)

Phone calls have been the hardest customer interaction to automate well, because callers interrupt, change topics, and expect an answer now rather than after a loading spinner. Google's Gemini 3.8 Live family, launched September 15, 2026, attacks that problem directly: native speech-to-speech models that keep a conversation going while performing tasks in the background. This article explains what Live is, what it costs per minute, what surrounds it in the Gemini lineup, and what it means for a service business that lives or dies by the phone.

Key takeaways

• Gemini 3.8 Live is a native speech-to-speech model that can perform tasks while keeping a conversation going, in standard and Extended Thinking versions.
• Published voice pricing is $0.005 per minute of audio input and $0.018 per minute of output, so a three-minute call costs roughly eight cents in model charges.
• Live Avatar adds real-time video presence for Gemini Enterprise; Flash TTS voices and Transcribe speech-to-text round out the voice stack.
• Live reaches developers through partners including Agora, LiveKit, Pipecat, and Vercel rather than as a finished phone product.
• For owners, the takeaway is better voice agents at lower cost, bought as configured systems measured on answered calls and bookings.

What Gemini 3.8 Live actually is

Most voice systems work in three relay legs: transcribe what the caller said, generate a text answer, then speak it aloud. Each handoff adds delay and loses tone, which is why older phone bots feel stilted. Gemini 3.8 Live is speech-to-speech natively, meaning it listens and speaks in audio directly while reasoning continues underneath. The practical difference is a conversation that tolerates interruptions, holds context across topic changes, and keeps talking while it looks something up.

Two versions launched together. The standard Live model suits responsive customer conversations. Live Extended Thinking trades some immediacy for deeper reasoning mid-dialogue, which fits calls where the agent must weigh options, such as triaging a repair request against parts availability and technician schedules. Both integrate through realtime partners including Agora, LiveKit, Pipecat, and Vercel, so developers can embed live voice in apps and phone flows without building audio infrastructure from scratch. For the reasoning engine behind text-heavy agent work, see our companion piece on Gemini 3.8 Flash for business.

One caution from experience with every voice launch: demo fluency is not deployment reliability. Accents, background noise, crosstalk, and callers who answer a question with another question are where voice agents earn or lose trust. Judge any Live-based system on recorded calls from businesses like yours, not on a staged conversation.

The per-minute math, worked honestly

Published pricing is refreshingly simple: $0.005 per minute of audio input (what the caller says) and $0.018 per minute of audio output (what the agent says). A typical three-minute service call with roughly even talk time costs about $0.0075 in input plus $0.027 in output, around four cents; a talkative agent or a five-minute call lands closer to eight to ten cents. Model audio is the cheapest line on the invoice.

The full per-call cost adds telephony minutes, the voice platform fee, and any CRM or booking integration charges, depending on the provider. Even stacked together, automated calls typically cost a fraction of staffed minutes, and the comparison that matters is against the silent alternative: the after-hours caller who hears ringing, leaves no message, and dials your competitor. If your phone is a revenue channel, price voice automation against recovered jobs, not against zero. Our guide to AI voice agents for small business walks through that calculation with realistic numbers.

For context on where this pricing sits in the market, the broader 2026 agent pricing comparison puts Live's per-minute rates next to per-token model pricing and platform subscriptions.

The supporting cast: Avatar, TTS, and Transcribe

Live does not stand alone; three sibling releases complete the voice stack. Gemini 3.8 Live with Live Avatar, announced September 24, 2026, adds real-time video presence with speech for Gemini Enterprise: a talking visual persona for live dialogue. Picture a video receptionist on a lobby tablet or guided on-screen help during a complex booking. It is enterprise-positioned today, so treat it as a direction signal rather than a purchase decision.

Gemini 3.8 Flash TTS and Flash-Lite TTS, from September 23, provide expressive text-to-speech for announcements, confirmations, and content playback where a full conversational model is unnecessary. Gemini 3.5 Transcribe, from August 2026, handles speech-to-text at 4.0 percent streaming word error rate across more than 85 languages, which matters for accurate call records and for businesses serving multilingual customers. Together they cover the whole loop: understand the caller, converse naturally, speak clearly, and write down what happened.

What this means for phone customer service

Three consequences follow for service businesses. First, the quality bar for automated calls rises: callers will compare every phone bot against fluid conversation, and stiff menu-driven systems will feel older by the month. Second, the cost floor falls: per-minute model rates this low make after-hours and overflow coverage affordable for shops that could never staff it. Third, the differentiator shifts from the voice itself to everything around it: the business rules, the calendar accuracy, the escalation judgment, and the record written to your system.

That third point decides vendor choices. A raw model does not know your services, prices, service area, or escalation policy. A configured AI voice agent wraps the model in exactly that knowledge plus call recording, summaries, and booking, and a tested voice receptionist workflow defines what happens when the caller is angry, vague, or facing an emergency. When evaluating any Live-powered offering, ask for the workflow documentation before the audio demo: the demo shows the model, but the workflow is what answers your phones in December.

How voice agents take calls today

While Live-based products mature, configured voice agents already handle the core phone jobs: answering every call in seconds, qualifying the need, booking straight to the calendar, taking messages with structured detail, and escalating emergencies by rule. The arrival of cheaper, more natural speech models improves the conversation layer of these systems without changing what makes them work: accurate business knowledge, tight calendar integration, and review of real transcripts. If missed calls are your problem now, waiting for the next model generation costs you this season's jobs. Deploy the configured system today on current models, and let better speech models upgrade the experience underneath you over time.

FAQ

What is Gemini 3.8 Live?

It is Google's native speech-to-speech model family, launched September 15, 2026, in standard and Extended Thinking versions. Unlike systems that transcribe, think, then speak in steps, it holds a live conversation and can perform tasks while the dialogue continues, which suits phone-style customer interactions.

How much does Gemini 3.8 Live cost?

Published pricing is $0.005 per minute of audio input and $0.018 per minute of audio output. A three-minute customer call therefore costs roughly eight cents in model audio charges before telephony and platform fees. Real deployments add those layers, so confirm the all-in per-call figure with your provider.

What is Gemini Live Avatar?

Announced September 24, 2026, Live Avatar adds real-time video presence with speech for Gemini Enterprise: a talking visual persona for live dialogue. Think reception-desk video or guided on-screen help rather than phone calls. Availability is enterprise-focused, so small businesses should treat it as a watch item.

Can Gemini Live replace my phone answering setup?

Not by itself. Live is a voice model, not a finished phone system: it needs telephony, calendars, business rules, escalation paths, and call records around it. The practical path is a configured voice agent built on this class of model, measured on answered calls and bookings.

Next step

Pull last month's call log and count the calls that arrived outside staffed hours; those are the minutes where per-minute voice pricing pays for itself fastest. To turn that count into a configured answer-every-call setup, start with the free six-step AI automation plan: start your AI automation plan. For the full service picture, see our AI voice agent page.