Gemini 3.8 Live Pricing (2026): What Voice Agents Cost

Short answer: Gemini 3.8 Live, released September 15, 2026, bills at $0.005 per minute of audio input and $0.018 per minute of audio output. A conversation with both sides talking runs about $0.023 per minute, or $1.35 per hour, and 1,000 three-minute calls work out to roughly $67.50 in model usage. For scale, that blended rate undercuts OpenAI's two closest published equivalents: its live-translation model at $0.034 per minute and its gpt-live-1 voice session at $0.05 per minute, with the caveat that OpenAI bills in a different shape and adds backend model usage on top.

Voice agents are now a line item, and the bill arrives in minutes, not tokens. Google shipped two new live models on September 15: Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Both were available in the Gemini API the same day they were announced, which makes this the week to price out what talking to a model actually costs. The short version: cheap enough for support lines and app copilots, expensive enough that a runaway loop of background calls shows up on the monthly invoice.

What Google shipped on September 15

Two models, one purpose. Gemini 3.8 Live is the default for low-latency voice agents: conversational, cheap at scale, and able to process visual input while the talk continues. Gemini 3.8 Live Extended Thinking adds background reasoning for harder tasks, narrating its progress while it works.

Google's announcement claims strong benchmark results. The Extended Thinking variant took the top spot on Artificial Analysis' Speech-to-Speech Quality Index at 82.6, and Google reports 68.6% on τ-Voice for agentic task completion. Those are vendor-reported numbers from launch day: directionally useful, worth re-testing on your own workflow.

The practical details:

• 97 languages, with mid-conversation switching.
• Tool and API calls run asynchronously while the model keeps talking, so a booking check does not freeze the call.
• Visual input is processed near real time.
• All generated audio carries a SynthID watermark.
• Developers get it in the Gemini API and AI Studio; consumers see it in Search Live and Gemini Live.

Gemini 3.8 Live pricing in per-minute terms

Google's rate card lists both token rates and per-minute equivalents, which is unusually convenient for voice work. Audio billing runs on 25 tokens per second, so a minute of streamed audio is 1,500 tokens, and $3.00 per million tokens lands at about $0.005 per minute.

Line item	Per 1M tokens	Per-minute equivalent
Audio input	$3.00	$0.005
Audio output (including thinking tokens)	$12.00	$0.018
Text input	$0.75	-
Text output	$4.50	-
Image / video input	$1.00	$0.002

Both 3.8 Live models sit on this one rate card. Billing follows token consumption, so dead air with the microphone closed costs nothing; a minute of overlapping speech costs more. The per-minute figures are there so you can estimate from wall-clock duration, which is how call centers already budget.

One number stands out. Output audio costs 3.6x input audio. In a typical call the model talks about as much as the caller, and that symmetry decides the bill: of the $67.50 that 1,000 three-minute calls cost, $54 is the model's own speech and $13.50 is listening. Four of every five dollars go to what the agent says.

The free tier is worth knowing about for testing, with one caveat. On the free tier, Google says content is used to improve its products. On the paid tier it is not. For a prototype that is fine; for customer calls, use paid.

What a voice agent costs per month

Pull the per-minute rates into the sizes teams actually run:

Workload	Total minutes	Cost at $0.023/min
500 support calls at 3 minutes	1,500	$33.75
1,000 support calls at 3 minutes	3,000	$67.50
10,000 minutes of a live assistant	10,000	$225.00

These are model-usage numbers only. Hosting, telephony, and the streaming platform in the middle are separate and not free. Tool use adds text tokens at $0.75 and $4.50 per million, and Google Search grounding runs 5,000 free searches a month across Gemini 3.x, then $14 per 1,000 queries.

How it compares with OpenAI's voice lineup

OpenAI publishes its voice pricing in a different shape, so the units matter. Rates below come from both vendors' pricing pages, read September 16, 2026.

Option	Published price	3,000 minutes/month
Gemini 3.8 Live	$0.005/min in + $0.018/min out	$67.50
Gemini 3.5 Live Translate	~$0.0368/min blended	~$110
Gemini 3.5 Transcribe Live	~$0.009/min blended	~$27
OpenAI gpt-live-1	$0.05/min session + backend usage	$150 + model tokens
OpenAI gpt-realtime-translate	$0.034/min	$102
OpenAI gpt-realtime-2.1-mini (audio)	$10/1M in, $20/1M out	depends on token mix

The shape differences are real. Google converts audio tokens to minutes at 25 tokens per second. OpenAI bills realtime models in tokens, and its gpt-live-1 session fee covers only the voice layer while model and tool usage is charged separately. Check both pages against recordings of your own traffic rather than trusting any single conversion.

The honest read: Gemini 3.8 Live lands as one of the cheapest ways to put a frontier voice model on a phone line. Google's own translation model, 3.5 Live Translate, costs more at an effective ~$0.0368 per minute, because translation is a narrower task on its own rate card.

What the Live API changes in your code

The Live API is a stateful WebSocket connection: audio in at 16kHz, audio out at 24kHz, barge-in so callers can interrupt, and real-time transcripts of both sides. Integration is either server-to-server or client-to-server with ephemeral tokens.

Most teams will not touch the WebSocket directly. Agora, LiveKit, Pipecat, Fishjam, Vercel, LangChain, and Vision Agents all wrap the Live API with the media plumbing handled, and Google lists all of them as launch partners. If you were on Gemini 3.1 Flash Live, now labeled legacy, the migration is mostly a model ID change, followed by a latency re-test.

Where this fits a BYOK stack

Voice is one leg of the stack. The Live API handles speech in, speech out, and transcripts; the reasoning about what the caller said can live anywhere, including a workspace where you've connected provider keys and switch models per task.

That split is where bring-your-own-key earns its keep. The voice leg bills at Google's rates above; the text leg bills at whatever model you route it to, from a $0.20-per-million budget tier to a frontier model, and you can change that choice without renegotiating anything with the voice vendor. How the workspace works covers the pattern, and there is a setup guide for connecting your Google key when you're ready.

For transcription alone, there is a cheaper lane: Gemini 3.5 Transcribe Live at roughly $0.009 per minute blended.

Which should you choose?

If you are building...	Best fit
A full-duplex voice agent at scale	Gemini 3.8 Live, $0.023/min blended
A voice agent that must reason mid-call	Gemini 3.8 Live Extended Thinking, same rate card
Live speech-to-speech translation	Gemini 3.5 Live Translate, ~$0.0368/min
Transcription and captioning only	Gemini 3.5 Transcribe Live, ~$0.009/min
On OpenAI's stack already	Compare gpt-realtime-2.1-mini at $10/$20 per 1M audio tokens against your actual token mix

What you give up

Three honest limits. This is an audio API, not a chat product: conversation history, key management, and multi-model routing stay your problem, and that integration work is a real cost beyond the per-minute rate. The benchmark claims are launch-day claims from Google and its leaderboard sources; independent evaluation is on you. And the SynthID watermark is baked into all output audio. It is inaudible, but verify your buyers are fine with it if you resell generated audio.

FAQ

Is Gemini 3.8 Live available on the free tier? Yes, both 3.8 Live models list a free tier for development, with the standard rate limits. Keep in mind that free-tier content is used to improve Google's products. The paid tier's terms are the ones to build a business on.

How does Gemini 3.8 Live compare with OpenAI's realtime API on price? The vendors bill differently, so there is no clean single number. Google's combined audio rate for 3.8 Live works out to about $0.023 per minute. OpenAI's gpt-realtime-2.1-mini bills $10 per million audio input tokens and $20 per million output tokens, with the per-minute cost depending on your token rate. OpenAI's flat $0.05 per minute gpt-live-1 session is the closest like-for-like voice fee, and backend model usage is extra.

Do I need a paid Google plan to use it in the API? No. The Gemini API is pay-as-you-go through Google AI Studio and Google Cloud billing. A consumer subscription like Google AI Pro is a separate product, not a requirement for API access. The overlap between API and subscription pricing is broken down in our Gemini API pricing comparison.

What happened to Gemini 3.1 Flash Live? Google's models page now labels it a legacy audio-to-audio preview and recommends moving to Gemini 3.8 Live. The pricing page currently groups it under the same rate card, but new work should target the 3.8 models.

Can I use Gemini 3.8 Live inside a regular chat app? Not in the same way as a text model. It connects over the Live API's WebSocket protocol for streaming audio, not the standard text-completion endpoints. The common pattern is a voice front end on the Live API with a text model behind it for anything that needs reading, searching, or long context, which is the kind of setup multi-model workspaces are built for.

One last budget note: the voice bill scales with minutes and the text bill with tokens, so the two age at different speeds. Doubling call volume doubles the voice line exactly; the reasoning layer is where routing choices still move the total, which is why the honest BYOK math tends to favor mixing models. Connect your Google key in the setup guide and the voice leg can start earning its keep.

---

Start your free trial → — 7 days, all providers, no credit card required.