01
Overview
Gemini 3.8 Live is designed for applications where speech is the interface rather than an afterthought. Google’s September 15, 2026 release adds stronger reasoning, asynchronous tool calls and live visual context to the Gemini Live API, making it relevant for support agents, booking systems, onboarding, sales qualification and multimodal assistants.
This guide focuses on the implementation decisions a developer needs to make. Google now lists stable model IDs gemini-3.8-live and gemini-3.8-live-extended-thinking. SDK syntax, quotas and operational limits should still be checked against the current Live API documentation before production deployment.
02
1. Choose between Gemini 3.8 Live and Extended Thinking
Start with the job your agent needs to complete.
Gemini 3.8 Live is positioned for scale, cost efficiency and fluid dialogue. Gemini 3.8 Live Extended Thinking is intended for harder multi-step tasks where deeper reasoning should happen without freezing the conversation.
For a simple appointment agent, FAQ assistant or guided workflow, standard Live may be sufficient. For a voice agent that must coordinate several tools, resolve ambiguous constraints or reason through a complex process, Extended Thinking is the more natural candidate to test.
Do not choose solely from benchmark rankings. Measure end-to-end task completion, latency, interruption handling and cost on your own workflow.
03
2. Prototype in Google AI Studio
Google says both new models are rolling out in Google AI Studio and through the Gemini API. AI Studio is the fastest place to validate conversational behavior before investing in a full media stack.
Test the kinds of interruptions, numbers, names, language switches and tool-triggering phrases your production users will actually produce. Voice agents often fail on edge cases that look trivial in text chat.
04
3. Build around the Live API
The Live API is the real-time transport layer for these models. A production architecture typically needs a client or telephony/media layer, a persistent Live API session, application state, tool definitions and backend services.
Google lists ecosystem integrations including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. These can reduce the work required to manage streaming media, but developers should verify which current Gemini 3.8 Live features each integration exposes.
05
4. Design asynchronous function calls
One of the key 3.8 Live capabilities is asynchronous function calling. The model can invoke tools or APIs in the background while continuing the spoken interaction.
That changes how a voice workflow should be designed. Instead of forcing the user to wait silently while a booking, CRM lookup or database request runs, the agent can acknowledge the request and continue naturally.
Your application still needs robust tool contracts. Validate arguments, handle timeouts, define retries and make failures understandable to the user. The model’s ability to keep speaking does not remove backend reliability requirements.
06
5. Add visual context only when it solves a real job
Gemini 3.8 Live can ground dialogue in visual input. That enables use cases such as troubleshooting a device, guiding a user through a screen, inspecting a physical object or answering questions about a live scene.
Visual input increases complexity and may affect bandwidth, privacy and cost. Do not stream video simply because the model supports it. Use it when visual context materially improves the task.
07
6. Handle multilingual conversations
Google says the models support 97+ languages and can automatically detect language changes. This creates an opportunity for global voice agents without building a separate speech stack for every supported language.
Production testing is still essential. Accent handling, domain vocabulary, names and mixed-language conversations can behave differently from benchmark conditions.
08
7. Use Extended Thinking selectively
Extended Thinking supports configurable thinking for more complex reasoning. For difficult workflows, it can reason in the background while continuing to respond or narrate progress.
That does not mean every request should use maximum reasoning. A production agent should reserve deeper reasoning for jobs that need it, especially when latency and cost matter.
09
8. Understand launch pricing
Google’s September 15 developer announcement lists audio pricing for the new Live models at $0.005 per minute of audio input and $0.018 per minute of audio output.
Treat those as launch-day figures and verify the current pricing page. Also account for your own media infrastructure, telephony, tool APIs, storage, logging and observability. Model audio pricing is only one part of the total cost of a production voice agent.
10
9. Test alphanumeric precision
Google specifically highlights improved handling of confirmation codes, claim numbers and technical data. If your workflow depends on these, build a test set around realistic noisy examples.
For high-stakes identifiers, repeat or display the captured value for confirmation instead of trusting a single spoken pass.
11
10. Plan for interruption and failure
A useful voice agent must handle users talking over it, changing their mind, correcting details and asking unrelated questions while a tool call is still running.
Test interrupted tool calls, stale results, duplicate actions and partial failures. Asynchronous execution makes conversations smoother, but it also creates more possible states for your application to manage.
12
11. Add observability before scale
Track latency, tool success, task completion, user corrections, abandonment and cost per successful task. A voice model can sound impressive while still failing the business process it is supposed to complete.
Avoid logging sensitive raw audio by default. Design retention, redaction and access policies appropriate to the application and jurisdiction.
13
12. Recheck operational limits
Before production launch, verify rate limits, session limits, regional availability, supported modalities and billing rules in Google’s documentation. The launch model IDs are the stable gemini-3.8-live and gemini-3.8-live-extended-thinking endpoints.
14
Practical workflow example
A booking agent might receive a spoken request, clarify date and party size, call an availability API asynchronously, keep the user informed while the request runs, present the options and then call the booking tool after explicit confirmation.
Extended Thinking becomes more relevant if the request includes several constraints or requires planning across multiple tools. Standard Live may be preferable when the workflow is straightforward and response speed is the priority.
15
FAQ
Do I need a separate speech-to-text model?
Google positions Gemini 3.8 Live as a native speech-to-speech Live model, so a separate cascaded transcription → text LLM → speech pipeline is not inherently required for the core interaction.
Can it call APIs while still talking?
Yes. Google documents asynchronous function calling for the new Live models.
Can I send visual context?
Yes. Google lists live visual context as a core capability.
Should I use Extended Thinking for every voice agent?
No. Use it when deeper multi-step reasoning is valuable. Standard Live is positioned for cost-efficient scale and fluid conversation.
What should I verify before launch?
Current model IDs, SDK/API syntax, pricing, rate/session limits, regional availability, integration support and privacy requirements.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.