Gemini 3.8 Live: What a Voice Agent for Member Services Costs to Run

Google made Gemini 3.8 Live generally available in the Gemini API on September 15, 2026, alongside a slower, deeper variant called Gemini 3.8 Live Extended Thinking. Both take a caller’s speech and answer in speech, in one model, and both can run tools in the background while they keep talking. Google’s launch post says the model switches among 97 languages mid-conversation.

Gemini 3.8 Flash TTS followed on September 22, and Gemini 3.5 Transcribe on August 26. Together they cover the live conversation, recorded speech and call records. This piece covers what each part does, the published rates, and what a call costs once you do the arithmetic. It also covers session limits, tool calling, phone lines, and what to test.

The short version: the per-minute rates on Google’s pricing page are a floor, not a per-call cost. The Live API bills the whole conversation again on every turn, so cost grows faster than call length. By our arithmetic, a five-minute call runs about 17 cents and a ten-minute call about 57 cents on published rates, before phone carrier charges. Sessions also end on their own unless you build for it. Budget for both.

What Google shipped for voice between August and September

Four model IDs matter for a voice agent, all listed in the Gemini API changelog:

The changelog describes gemini-3.8-live as the "default option for most low-latency voice agent experiences." It calls the Extended Thinking model a "high-reasoning audio-to-audio model."

Availability differs by surface. Google’s launch post lists both Live models in the Gemini API and Google AI Studio. Gemini Enterprise has them in private preview. Gemini Enterprise for Customer Experience is listed as "coming soon." This post covers the Gemini API, where agencies are most likely to build.

Older voice models now point here. The deprecations page names gemini-3.8-live as the replacement for gemini-3.1-flash-live-preview and gemini-2.5-flash-native-audio-preview-12-2025. Neither has a shutdown date as of late September. If you track dates like these, our list of Gemini shutdown dates this fall covers the other retiring IDs.

What Gemini 3.8 Live does in a call

The model takes audio, text, images and video in, and returns audio and text. Its model card lists a 131,072-token input limit and a 65,536-token output limit.

Language handling is automatic. Google’s post says the model "automatically detects and transitions between 97 supported languages mid-conversation." The Live API capabilities guide adds that native audio models choose the language themselves and "don’t support explicitly setting the language code." You can narrow it through the system instructions. For a helpline that serves callers in English and Spanish, that matters: the caller does not press a button to switch.

Three behaviors changed from earlier Live models, per the gemini-3.8-live model card:

• Proactive audio is always on. The model can decide not to respond to speech that is not meant for it. Setting it to false returns an error.
• Affective dialog is gone. The card says it “is removed from the API.” Code that set it will need editing.
• No thinking level setting. The card lists thinking as “interleaved reasoning” and says `thinking_level` is not supported.

When a caller talks over the agent, the server flags the turn as interrupted, and your client must drop queued audio. Google’s post says all audio its AI products generate carries a SynthID watermark.

Standard or Extended Thinking

Most member services calls are lookups and routing: renewal dates, event times, where to send a form. Standard Gemini 3.8 Live fits that work.

Extended Thinking is for calls that need more reasoning, such as sorting out a billing question across several records. Google’s post says it "reasons and speaks simultaneously." The post does not say how much latency it adds. Test both on your own call scripts.

One difference will shape your code. The Extended Thinking model card lists function calling as "Supported (Async only)." Every tool it calls runs in the background. The standard model defaults to background calls too, but the card does not describe it as the only mode.

Both models share one price line, covered next. Thinking tokens bill as output, so a model that reasons more costs more per call.

The Gemini 3.8 Live rate card

The launch post gives no prices. The Gemini API pricing page groups Gemini 3.8 Live, 3.8 Live Extended Thinking and 3.1 Flash Live Preview under one set of paid-tier rates, as of late September 2026:

The page says audio runs at 25 tokens per second. Grounding with Google Search includes 5,000 free requests a month, shared across Gemini 3.x models, then $14 per 1,000.

The free tier costs nothing, but the page marks free-tier data as used to improve Google’s products. The paid tier is marked "No." For a helpline, that settles which tier real calls belong on.

Unlike the Flash TTS rows on the same page, the Live rows show no scheduled price change for January 1, 2027. Our look at Gemini 3.8 Flash’s cost per task explains why the rate card and the bill often diverge.

Why a Gemini 3.8 Live call costs more than its minutes

The per-minute rates suggest a five-minute call costs about six cents. That figure leaves out how the Live API bills context.

Google’s Live API best practices page states it directly: "The API charges you per turn for all tokens present in the session context window." It adds that "Past tokens are re-processed and accounted for in each new turn." Accumulated audio bills at the audio input rate every time.

So the tenth answer in a call pays again for everything said in the first nine. Cost tracks the number of turns and the length of the call together. A Google staff member confirmed the same in a July 2026 developer forum reply, saying the model keeps the raw audio from past turns rather than converting it to text.

Transcription adds a small charge. If you turn on input or output transcripts, the text bills at the $4.50 output rate on top of the audio.

A worked estimate for a five-minute member call

Assumptions: a five-minute audio-only call with 10 exchanges. Each exchange has 15 seconds of caller speech and 15 seconds of agent speech. Both sides stay in context. A 2,000-token system prompt and tool list rides along on every turn. Audio counts at 25 tokens per second.

That is about 3.5 cents a minute, three times the six-cent floor. At 1,000 such calls a month, the model line comes to about $172.50.

Length makes it worse. Run the same pattern for ten minutes and 20 exchanges, and the total reaches about $0.57. The floor for that call is about $0.115. Twice the length costs more than three times as much.

Three costs sit outside this table. Phone carrier charges come on top, and we have not priced them. Extended Thinking adds thinking tokens at the output rate. Search grounding past the free allowance adds $14 per 1,000 requests.

Context window compression limits the growth on long calls. Google’s example sets compression to trigger at 25,000 tokens with an 8,000-token sliding window. By the 25-tokens-per-second rate, 25,000 tokens is about 16 minutes of audio, so most member calls would never reach it.

Session limits that shape a helpline design

The Live API has two clocks, and both end calls if you ignore them. The session management guide sets these limits:

• Session length: audio-only sessions stop at 15 minutes without context window compression. Audio with video stops at 2 minutes. With compression turned on, sessions can run without a time limit.
• Connection length: the WebSocket connection itself lasts “around 10 minutes.” When it ends, the session ends too.
• Warning: the server sends a GoAway message with a `timeLeft` field before it cuts the connection.
• Resumption: a resumption token reconnects to the same session for 2 hours after the last session ends.

Any Gemini 3.8 Live call that might pass ten minutes needs session resumption built in from day one. A benefits question can easily run that long, and the caller should hear nothing when the connection swaps.

Google’s rate limits page publishes no numbers for the Live models. It says limits "can be viewed in Google AI Studio." Check concurrent session limits there before a newsletter sends a thousand members to one phone number.

Tool calling now runs in the background

A member services agent needs to look things up in your AMS, CRM or event system. In the Live API, that means function calling.

The gemini-3.8-live model card says "Async execution (behavior: NON_BLOCKING) is now the default function calling mode." The model can say "Let me pull up your record" and keep talking while your lookup runs. Google’s post describes it as executing "tools and API calls in the background while continuing the conversation."

When the result returns, your code tells the model how to use it. The Live API tools guide lists three options: INTERRUPT delivers it at once, WHEN_IDLE waits for the model to finish speaking, and SILENT stores it for later.

Two details catch teams out:

• You handle every tool response. The tools guide says the Live API “doesn’t support automatic tool response handling.” Your code receives the call, runs it and sends the result back.
• The guide lags the model cards. As of late September, the Live API tools page still shows its support table for the older 3.1 and 2.5 Live models. Go by the 3.8 model cards, and confirm in testing.

Google Search grounding works on both 3.8 Live models. Code execution, structured outputs, caching and the Batch API do not, per the model cards.

Getting calls in from a phone line

The Live API speaks WebSockets, not phone lines. Something has to sit between a phone number and the model. Google does document a path, though not as a built-in feature.

The Google AI account on DEV published a telephony tutorial in April 2026. It bridges Twilio Media Streams to the Live API through a small FastAPI server on Google Cloud Run. The bridge converts the phone line’s 8 kHz mu-law audio to 16 kHz PCM for the model, then converts the model’s 24 kHz output back down. The tutorial used gemini-3.1-flash-live-preview, so expect to swap in the 3.8 model ID.

It carries a security warning worth repeating. Its outbound-call endpoint is unauthenticated, and the author says it must be secured in production.

The launch post also names developer platforms that support Gemini 3.8 Live, including LiveKit, Pipecat, Agora and Vercel. We have not tested any of them for this post.

For a website voice widget instead of a phone number, the Live API guide supports a browser connecting directly. It suggests ephemeral tokens so your API key never reaches the browser.

Where Flash TTS and 3.5 Transcribe fit

Gemini 3.8 Flash TTS turns text into speech. The speech generation guide lists 30 prebuilt voices, over 130 languages for Flash TTS and over 100 for Flash-Lite TTS. It can output 8-bit G.711 mu-law or A-law audio at 8,000 Hz, the format phone systems use. That suits greetings, hold messages and event-day announcements. Paid rates are $0.50 per million text input tokens and $9.00 per million audio output tokens through December 31, 2026, doubling on January 1, 2027. Flash-Lite TTS output is $6.00, also doubling.

Gemini 3.5 Transcribe turns speech into text. Its model card lists 85 or more languages, files up to one hour, and speaker labels for up to eight speakers, with three or more marked experimental. Word timestamps and speaker labels cap files at 30 minutes. The streaming version, gemini-3.5-transcribe-live, runs 10 minutes per session. File transcription is priced at $0.003 per minute of audio in and $0.002 per minute of text out.

Note the language gap. Gemini 3.8 Live handles 97 languages by Google’s count; Transcribe lists 85 or more. A caller the agent understands may not get a clean record. Our post on transcription model retirement dates for captioning covers what caption teams should check when switching transcription models.

What to test before a real caller hears Gemini 3.8 Live

A voice agent fails in ways a chatbot does not. Run these checks on the paid tier with your own scripts:

1. Real call length. Run calls past 10 and 15 minutes. Confirm resumption works and the caller hears no gap.
2. Real cost. Log token usage per call and compare it to the worked estimate above. Check whether silence your bridge streams is being billed.
3. Lookups that return late. Slow down your AMS on purpose. Confirm the agent waits for the answer rather than guessing a renewal date.
4. Language switching. Have a tester switch languages mid-call. Check that transcripts and handoff notes still make sense.
5. Noise and interruptions. Test from a car, a conference hall and a speakerphone. Talk over the agent and confirm queued audio stops.
6. Handoff to a person. Define when the agent transfers, and test that it does.
7. Standard versus Extended Thinking. Time both on your hardest scripts, and compare thinking-token cost.

For comparison, our July coverage of xAI’s no-code voice agent builder shows a no-code approach to the same job.

Frequently Asked Questions