Gemini 3.8 Live: What It Costs and How to Build Real-Time Voice Agents

Published 23 September 2026 by Jake Hissitt, Founder of Stob.AI.

Most model comparisons treat every model as a request-response API. Gemini 3.8 Live is not that. It is a streaming model built for continuous voice and video conversation, stable since 15 September 2026, and it is the only model in our comparison tool designed for that shape of work.

If you are building a phone agent, a voice concierge or a screen-sharing assistant, this is the relevant page.

Pricing, and why it is different

The per-minute figures are the ones to budget with. A ten-minute call with roughly balanced speaking time costs in the region of 12 cents in model spend. Telephony, transcription fallbacks and your own infrastructure sit on top.

There is no prompt caching on Live. A long system prompt is paid for on every session, so keep instructions tight and push reference material into tools rather than the prompt.

What it supports, and what it does not

• Supported: async function calling, search grounding, 131,072 token context, up to 65,536 tokens of output, streaming audio and video input.
• Not supported: code execution, structured outputs, prompt caching.

The missing structured outputs matter. If your voice agent needs to write clean records into a CRM, do not rely on the live model to emit perfect JSON mid-conversation. Have it call a function with named arguments, or hand the transcript to a cheap text model such as GPT-6 Luna at the end of the call and let that produce the record.

The Extended Thinking variant

Gemini 3.8 Live Extended Thinking runs at the same published rates and adds background reasoning while the conversation continues, with async-only tool calling. Use it where the agent must work something out without an awkward silence — checking eligibility, reconciling conflicting information, planning a multi-step action — and use standard Live where snappy turn-taking matters more than depth.

A workable architecture

1. Live model handles the conversation: listening, speaking, interruption handling.
2. Functions expose a small, tightly scoped set of actions: look up a customer, check availability, create a callback. Validate every argument before it touches a real system.
3. A cheap text model processes the transcript after the call into a structured summary, a CRM record and a next action.
4. Guardrails at the point of action. The live model proposes; your own code checks current state and permissions before anything is written or sent.

That last point is not optional. A voice agent is an untrusted-input surface by definition. The separation of interpretation from execution is the same principle we set out in 7 things before building a reliable AI system.

Frequently asked questions

How much does Gemini 3.8 Live cost per minute?

Roughly $0.005 per minute of audio input and $0.018 per minute of audio output, based on Google's published token rates of $3 and $12 per million audio tokens.

Does Gemini 3.8 Live support prompt caching?

No. Every session pays full price for the system prompt, so keep it short.

Can it return structured JSON?

Structured outputs are not supported. Use function calling during the call, and a separate text model afterwards to produce records.

What is the difference between Live and Live Extended Thinking?

Extended Thinking adds background reasoning during the conversation with async-only tool calling, at the same published rates. Standard Live is tuned for lower-latency turn-taking.

What context window does it have?

131,072 tokens, with up to 65,536 tokens of output — much smaller than the million-token text models, which is why session state management matters.

What to do next

Prototype one narrow call type end to end — booking confirmations, say — with a hard limit on what the agent may action without a human. Compare the full Gemini line in Best Gemini Model 2026, or every model side by side in the AI model comparison tool.