Gemini Live API Pricing (2026): Per-Minute Costs by Model

Gemini Live API pricing starts at $0.005 a minute for audio input and $0.018 a minute for audio output on Gemini 3.8 Live, Google's default real-time voice model. Text, image, and video input cost less, and two specialist models (Transcribe Live and Live Translate) run about $0.009 and $0.037 a minute.

Those per-minute rates assume each second of audio bills once. A real call bills a lot more, because the Live API charges every turn for the whole conversation so far.

On an illustrative five-minute support call, Gemini 3.8 Live comes to about $0.03 a minute.

If you stretch the call to 15 minutes, it's closer to $0.07 a minute, which lines up with the $0.070 a minute Cekura’s benchmark worked out for Gemini 3.1 Flash Live at Google’s published rates on live test calls.

There's also a deadline. Two of Google’s Live models, Gemini 3.1 Flash Live Preview and the Gemini 2.5 Flash Native Audio preview, shut down in the Gemini Developer API on November 17, 2026, and the replacement bills listening time differently.

We went through Google's rate card, its Live API billing rules, and its deprecation schedule, then priced calls turn by turn, so you can see what each model costs per minute, per call, and per month, and why your invoice won't match the rate card.

Pricing verified against Google's pricing pages on 28 October 2026. Verify with the vendor before you commit.

Gemini Live API Pricing at a Glance

Google's Gemini Developer API pricing page lists six models that run on the Live API. This is the paid-tier rate card, per 1 million tokens (Google's own per-minute equivalents in brackets):

A few things stand out before you get into any one model:

• The three conversational Gemini 3 models share one rate card. Moving between 3.8 Live, 3.8 Live Extended Thinking, and 3.1 Flash Live leaves every per-token price where it is. What changes is how many tokens you burn.

• Output costs four times input on audio, and text output covers the model's thinking too.

• Every model on the list has a free tier, with the catch covered further down.

Gemini Live API Pricing by Model

Each Live model below gets its own rate card, plus what it's for and what makes its bill move. All prices are paid-tier rates per 1 million tokens.

Gemini 3.8 Live

Gemini 3.8 Live is the model Google recommends for "most low-latency voice agent experiences," and the one both of the retiring models point to.

• Best for: Production voice agents that need fast replies and asynchronous function calls

• Watch for: Proactive audio is permanently on, so Google bills input for the whole time the session is listening. The model page also lists caching as "Not supported," so there's no discount on re-sent conversation history.

• Thinking: Interleaved reasoning is built in, and you can't set a thinking level. Whatever it reasons through bills as text output.

Gemini 3.8 Live Extended Thinking

Same price list as 3.8 Live, with configurable "background reasoning" for calls where the agent has to work something out while it keeps talking.

• Best for: Calls with eligibility checks, multi-step routing, or policy lookups

• Watch for: Thinking levels are low, medium, and high (minimal isn't supported). Higher levels mean more output tokens at $4.50 per million, so on the same call it can cost more than 3.8 Live even though the rate card is identical.

• Tool calls: Asynchronous only, so your agent keeps talking while a function runs

Gemini 3.1 Flash Live Preview

Google now calls this a "legacy preview model," and it shuts down on November 17, 2026.

• Best for: Nothing new. Plan your move to 3.8 Live now.

• Watch for: Proactive audio isn't supported here, so Google bills audio only while you're streaming input. That’s the biggest billing change when you migrate, and the shutdown section below covers the rest.

• Thinking: Defaults to minimal for the lowest latency, with low, medium, and high available

Gemini 2.5 Flash Native Audio

The older native-audio model (gemini-2.5-flash-native-audio-preview-12-2025 in the Developer API) has cheaper text and pricier video than the Gemini 3 cards.

• Best for: Existing deployments that haven't migrated yet

• Watch for: The Developer API preview shuts down on November 17, 2026. On Google's enterprise platform, a generally available version (gemini-live-2.5-flash-native-audio) is still listed at the same rates, although it’s worth noting that this too will retire in December 2026.

Gemini 3.5 Transcribe Live

A speech-to-text model that streams over the Live API. It returns text and doesn't hold a conversation.

• Best for: Live captions, call transcripts, and the speech-to-text layer of a chained voice agent

• Watch for: Google's blended rate assumes 175 text tokens per minute of speech. Fast talkers produce more.

Gemini 3.5 Live Translate

A preview speech-to-speech translation model that supports 70+ languages.

• Best for: Real-time interpretation between a caller and an agent or a person

• Watch for: Billing counts total input and output audio at 25 tokens per second, so a minute of translated speech in both directions is close to that $0.0368 figure

How the Gemini Live API Bills a Session

Google's Live API best-practices guide explains how many tokens a session burns, which is where the difference between the brochure rate and your invoice comes from.

Audio Counts as 25 Tokens a Second, Both Ways

Caller audio and model audio both count as 25 tokens per second, or 1,500 tokens a minute. At $3 and $12 per million, that works out to $0.0045 a minute in and $0.018 out, which Google rounds to $0.005 and $0.018.

Every Turn Re-Bills the Whole Session

Google charges per turn for all tokens in the session context window, and a turn is one user input plus the model's reply. Every earlier caller utterance and every earlier agent reply gets re-processed and re-billed, on every new turn.

The Live API keeps that history as raw audio tokens, and Google's guide says it bills them "at the standard audio input rate on every turn." Your system instructions ride along too, as text input.

No Caching Discount on the History

OpenAI's Realtime API softens the same re-billing with prompt caching. Gemini 3.8 Live lists caching as unsupported, so turn 20 pays full audio input price for everything said in turns 1 through 19. That one difference does more to shape long-call costs than any line on the rate card.

Thinking Bills as Text Output

Google's output price is labeled "including thinking tokens." Gemini 3.8 Live reasons on its own, and Extended Thinking lets you push the level up.

Either way, the reasoning bills at $4.50 per million text output tokens, and Google's Agent Platform pricing page points to the thoughts field in the response metadata if you want to audit it.

Proactive Audio Bills the Listening Time

With proactive audio on, the model decides when to respond, and Google bills input tokens the entire time the Live API is listening. Output only bills when it speaks.

On Gemini 3.8 Live and 3.8 Live Extended Thinking, proactive audio is permanently enabled. On Gemini 3.1 Flash Live, it isn't supported, and you pay for audio only while you're streaming it.

What a Gemini Live Call Costs: Worked Examples

Take an illustrative five-minute support call on Gemini 3.8 Live, priced with Google's published token rules:

• 10 turns: the caller speaks for 10 seconds (250 audio tokens), the agent replies for 15 seconds (375 audio tokens)

• 2,000-token system prompt with tool definitions, re-billed every turn as text input

• No caching, since 3.8 Live doesn't support it

• Thinking tokens, proactive-audio listening time, and transcripts left out, so treat these totals as a floor

A single turn grows like this over the call:

By turn 10, each turn costs 3.5 times as much as turn 1, even though every caller question and every agent answer is the same length.

Add up all 10 turns and stretch the call to 15 minutes (30 turns of the same shape), and you get:

So the five-minute call runs about three times the rate-card math, and the 15-minute call about six and a half times. Gemini 2.5 Flash Native Audio lands within a few percent of these numbers (its cheaper text helps a little), since audio is the same price.

Per 1,000 Calls and Per Month

Based on the five-minute call above ($0.15 a call):

• 1,000 calls: about $152

• 10,000 calls a month (50,000 minutes): about $1,519

• 50,000 calls a month (250,000 minutes): about $7,594

At the 15-minute call ($1.02 a call), 1,000 calls cost about $1,018, which is nearly seven times as much for three times the talk time.

None of this includes telephony or an orchestration platform, which add their own per-minute fees.

Our Retell AI pricing breakdown shows what that platform layer adds on top of the model, and if you're setting your own per-minute price for customers, our guide on how to price AI voice agents walks through the models that hold up.

Why Real Bills Run Above the Per-Minute Math

If your invoice says your agent costs more than $0.005 plus $0.018 a minute, nothing's broken. These are the five places the extra comes from, roughly in order of size.

1. The Conversation History Re-Bills Every Turn

It’s the largest item, and it grows with every turn: three times the rate-card math on a five-minute call, six and a half times on a 15-minute one. A long system prompt makes the curve steeper, and with no caching on 3.8 Live, there’s no discount to claw it back.

Cekura's speech-to-speech benchmark shows the same effect on real calls. It priced Gemini 3.1 Flash Live Preview at $0.070 a minute at published rates across 246 live test calls.

That's right next to our 15-minute figure, and about three times the $0.023 list price for a full minute in plus a full minute out, a stricter baseline than the worked example’s, which counts only the seconds each side talks.

2. Thinking Tokens Bill as Text Output

Gemini 3.8 Live always reasons, and Extended Thinking at a high level reasons more. Every thought bills at $4.50 per million text output tokens, and none of it shows up in a per-minute audio rate.

3. Proactive Audio Bills While the Model Listens

On 3.8 Live, input bills for the whole time the session listens. In the five-minute example, the caller talks for 100 seconds.

If the other 200 seconds of listening bill as input too, that's roughly $0.015 more per call before any of it is re-billed in later turns. Google doesn't say how much of that listening audio stays in the context window, so measure it on your own traffic.

4. Transcripts Bill at the Text Output Rate

When you turn on input or output transcription, Google charges every transcript token at the text output rate, on top of the audio.

Using Google’s own estimate of 175 text tokens per minute of speech, transcribing both sides of the five-minute call adds only about $0.003.

5. Search Grounding and Video Frames Bill Separately

Grounding with Google Search is free for the first 5,000 search requests a month, shared across all Gemini 3 and newer models, then costs $14 per 1,000 requests, up to 1,500 requests a day.

Video input adds $0.002 a minute on the Gemini 3 cards, but audio plus video sessions stop at two minutes unless you turn on context window compression.

How to Keep a Gemini Live Bill Down

• Turn on context window compression. Set a trigger and a sliding window (Google's example is 25,000 and 8,000 tokens), and the API will drop older tokens once it reaches the threshold, then bill only the retained history.

Audio-only sessions stop at 15 minutes without it anyway, according to Google's session management guide. On a 30-minute call of our example shape, those settings bring the cost to about $2.69.

Without compression the session would have stopped at 15 minutes, and if it hadn’t, the full history would have pushed it to about $3.72.

• Keep the system prompt short. It bills on every turn. Move product catalogs and policy documents behind a function call.

• Use transcription only where you need it. Caller-side transcripts for QA are often enough.

• Pick the reasoning you need. Extended Thinking at high is for calls that need it, and plain 3.8 Live is cheaper for everything else.

Two Gemini Live Models Shut Down on November 17, 2026

Google's deprecations page lists the same shutdown date and the same replacement for both:

Moving from 3.1 Flash Live to 3.8 Live looks free on paper, and three things change underneath it:

• Listening time starts billing. 3.1 Flash Live bills audio only while you stream input. 3.8 Live has proactive audio on permanently.

• Your thinking_level setting has to go. 3.8 Live doesn't support it, and Google says to omit it from setup, so you lose the minimal level that kept 3.1 Flash Live's reasoning light.

• Tool calls go asynchronous by default. 3.1 Flash Live ran function calls synchronously, so the model waited for each tool response before it replied, and your agent logic may need changes.

From 2.5 Flash Native Audio, audio stays the same price, and text gets more expensive. On a voice agent, audio accounts for most of the bill, so the move is close to neutral unless your agent leans hard on text input or thinking.

The bigger question is quality, and the benchmark section below shows that 3.8 Live doesn't pass as many of Cekura's test scenarios as 3.1 Flash Live.

Is the Gemini Live API Free?

These are two separate questions, and people searching for them often mix up the app and the API.

Is the Gemini API Still Free?

Yes, the Gemini API still has a free tier, and every Live model is "free of charge" on it, up to rate limits that Google shows in AI Studio. The catch is that on the free tier, Google's pricing page marks your content as used to improve its products. On the paid tier, it doesn’t.

To move to paid, Google's billing guide says you link a billing account and prepay at least $5 in credits. Tier 1 accounts have a $250 billing cap, which rises to $2,000 at Tier 2 (after $100 paid and three days) and $20,000 to $100,000+ at Tier 3.

For a production voice agent handling customer calls, the free tier's data terms rule it out. Use it for prototyping.

Is Gemini Live Free to Use?

Yes, Gemini Live in the consumer Gemini app is free with a Google Account. Google's plans page lists it in the $0 Free plan, and paid plans raise the usage limits:

• Google AI Plus: $4.99 a month, 2x the Free plan's usage limits

• Google AI Pro: $19.99 a month, 4x the Free plan's usage limits

• Google AI Ultra: $99.99 a month for 5x Pro's limits, or $199.99 for 20x

None of these subscriptions include Live API usage. The app and the API bill separately, so a Pro subscription won't pay for your voice agent's tokens.

Gemini Developer API vs Gemini Enterprise Agent Platform Pricing

Google sells the Live API two ways: the Gemini Developer API (keys from AI Studio) and Gemini Enterprise Agent Platform (the platform formerly called Vertex AI).

For Gemini 3.8 Live, the Agent Platform pricing page lists the same per-token rates. The differences are in what you get around them:

6,192 tokens a second at $1 per million comes to about $0.37 for every minute the avatar speaks, roughly 20 times the cost of the audio itself. Google doesn't bill avatar video while it's listening.

For most teams, the choice comes down to the free tier and fast setup on the Developer API, against Cloud billing, regional endpoints, avatars, and the generally available 2.5 model on Agent Platform (which will retire in December 2026).

The token price is the same either way. The Agent Platform's Live API overview lists its supported models.

Gemini Live vs OpenAI Realtime and Grok: Cost per Successful Call

A fair comparison weighs price per minute against how often the agent gets the call right.

List Price on the Same Call

Here's our illustrative call priced on Gemini 3.8 Live and on OpenAI's Realtime models, using OpenAI's rates as of 26 October 2026 and the same 10-second caller turns, 15-second agent replies, and 2,000-token prompt.

The OpenAI figures come from our OpenAI Realtime API pricing breakdown:

On a short call, Gemini 3.8 Live costs about half what OpenAI's flagship does with caching working.

The difference is smaller on a long call. 15 minutes on 3.8 Live costs about the same as gpt-realtime-2.1 with caching, and three times what the cached mini costs. The lack of caching drives most of that.

What Cekura's Benchmark Shows

Cekura's speech-to-speech benchmark (last updated October 22, 2026) ran 12 realtime models through the same Pipecat agent, prompt, and tools on 82 caller scenarios, three times each, covering clinic booking calls and Medicare eligibility calls.

A call passes only when the agent says the right things and saves the right data.

The benchmark lists no verified per-minute rate for the two Gemini 3.8 models, so they're left out of its cost ranking. Their rate card matches 3.1 Flash Live's, so expect a similar per-minute figure before proactive audio and thinking.

Where Each Gemini Model Wins and Loses

Gemini 3.1 Flash Live was reliable on bookings (93.2% success) and the slowest to reply in the field, at 2.88 seconds. Its cost per successful call, $0.177, beat gpt-realtime 2.1's $0.191, which makes it a reasonable value pick, and it's the one Google is retiring.

Gemini 3.8 Live replies faster (1.91 seconds, 5th of 12) and ranked last on reliability. The Gemini 3.8 Live benchmark page shows where it struggled most: calls where the caller corrects themselves or changes their mind passed 38% of the time, against a median of 81% for the field.

Gemini 3.8 Live Extended Thinking posted one of the strongest Medicare results among the realtime models, 69.6% success, against 52.2% for plain 3.8 Live and 31.9% for 3.1 Flash Live. The extra reasoning costs more in output tokens, and on eligibility calls it bought a much higher pass rate.

The top of the field pulls well ahead. In the Gemini 3.8 Live vs Grok Voice 3.0 comparison, Grok wins six of eight measures and ties on interruptions. You can line up any two models in the S2S benchmark comparison tool.

Which Option Fits Your Agent

Gemini 3.8 Live makes sense if you:

• Run short, scripted calls where a fast reply counts for more than handling a caller who changes their mind

• Want the lowest list price on short calls without depending on a cache

• Already build on Google Cloud and want one bill

Pick Gemini 3.8 Live Extended Thinking if you:

• Handle eligibility, compliance, or routing calls where a wrong record costs you the most

• Can absorb a little more reply time and some thinking tokens per turn

Look past Gemini Live if you:

• Need the highest pass rate on every run, where Phonic 1, Grok Voice 3.0, and gpt-realtime 2.1 lead the benchmark

• Run long calls, where a cached OpenAI model can cost the same or less

• Want the lowest cost per successful call, where gpt-realtime 2.1 Mini leads at $0.066

Test the November Migration Before Your Callers Do

The Gemini Live API pricing card barely moves on November 18. 3.1 Flash Live users keep the same per-token prices but start paying for listening time, and 2.5 Flash Native Audio users pay more for text.

The bigger change is the model behind the price, and the benchmark shows the replacement handles a different set of calls well.

Finding out whether 3.8 Live handles your callers, at what cost per successful call, takes hundreds of calls run several times each. That's the work Cekura automates, before a model switch reaches production and after it does.

With Cekura, teams can:

• Pre-production testing: run the same caller scenarios against 3.1 Flash Live, 3.8 Live, Extended Thinking, and any other model, repeat each one across multiple runs, and compare pass rates and saved-data accuracy side by side

• Infrastructure testing: check how each model handles interruptions, background noise, long pauses, and callers who correct themselves, the conditions where Cekura's benchmark saw 3.8 Live slip

• Observability: replay production calls with the original audio against the new model before you switch, then monitor live calls for drop-offs and missed workflow steps after you do

Native integrations cover Retell, Vapi, ElevenLabs, LiveKit, Pipecat, and Bland, and Cekura is SOC 2-, HIPAA-, and GDPR-compliant for transcript redaction, role-based access, and audit trails.

Book a demo to see how 3.8 Live performs on your own agent’s calls before you switch.

Frequently Asked Questions

How Much Does the Gemini Live API Cost per Minute?

The Gemini Live API costs $0.005 a minute for audio input and $0.018 a minute for audio output on Gemini 3.8 Live, before conversation history is re-billed.

On an illustrative five-minute call with 10 turns, the real cost comes to about $0.03 a minute, and on a 15-minute call about $0.07 a minute, because every turn re-bills the whole session.

Is the Gemini Live API Cheaper Than OpenAI's Realtime API?

Yes, on short calls the Gemini Live API is cheaper than OpenAI's flagship Realtime model, and on long calls the difference shrinks.

An illustrative five-minute call costs about $0.15 on Gemini 3.8 Live against $0.33 on gpt-realtime-2.1 with caching. At 15 minutes, both cost about $1 a call, because OpenAI caches conversation history and Gemini 3.8 Live doesn't.

Does the Gemini Live API Charge for Silence?

Yes, on Gemini 3.8 Live, the API charges for input audio the whole time it's listening, silence included.

Proactive audio is permanently on in 3.8 Live and 3.8 Live Extended Thinking, and Google bills input tokens while the model listens. Gemini 3.1 Flash Live doesn't support proactive audio and bills audio only while you stream it.

Does the Gemini Live API Support Context Caching?

No, Gemini 3.8 Live doesn't support context caching. Google's model page lists caching as not supported, so earlier turns re-bill at the full $3 per million audio input rate on every new turn.

Context window compression is the main lever for long calls, since it drops older tokens once the session passes a threshold you set.

How Long Can a Gemini Live API Session Last?

A Gemini Live API session lasts up to 15 minutes for audio only and 2 minutes for audio plus video, unless you enable context window compression.

With compression, sessions can run for an unlimited time. Each WebSocket connection lasts around 10 minutes, and session resumption tokens, valid for two hours, let a session continue across connections.