Gemini 3.8 Live gives businesses a new way to build voice agents that keep talking while connected tasks run. The useful question is whether a conversation ends with the right result in your business system. A fluent answer alone cannot establish that.
Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on 15 September 2026; its announcement was updated on 17 September. This report checks the official release and developer documentation as of 22 September 2026, then sets out a practical evaluation plan. It is a research-based explanation, not a claim that we have tested either model.
What Google announced
The official Gemini 3.8 Live announcement describes two models: a version focused on efficient live dialogue and an Extended Thinking version for more complex work. Both combine conversation with visual context and background tasks.
That matters when someone needs to correct a request while a system is still checking information. Instead of treating the call as a neat question followed by a finished answer, the application has to manage a conversation and work happening at the same time.
For a small business, the first useful application might be an internal appointment assistant, a staff training simulation or a guided product-information lookup. Those are possible pilot designs, not demonstrated results for these models. Start with the bounded-use principles in our guide to using AI in a small business without losing control.
Check the access route before planning a rollout
Google's launch page says the models are rolling out through the Gemini API and Google AI Studio. It separately describes enterprise private previews, upcoming enterprise integrations and different routes into Gemini Live, Search Live and Workspace. That is not a promise that every business account has every feature.
Write down which route you intend to use: a custom application built with the API, a feature in a consumer subscription, or a managed business service. Then check the actual model and feature available to that account. A subscription feature and an API deployment have different setup, administration and billing decisions.
Keep this access check small and concrete. Ask the person implementing the pilot to record the model identifier, service, account type, supported location and any preview restriction visible in the relevant service. Do not build a customer promise around an announcement until the required route is available to you.
Choose the model around the task
Google's Gemini 3.8 Live model documentation positions the standard model for responsive voice interactions. Its stable identifier is gemini-3.8-live. The page also documents interleaved reasoning and asynchronous function calling: a connected operation can run without making the entire conversation wait.
The Extended Thinking documentation describes background reasoning for more involved problems and uses the identifier gemini-3.8-live-extended-thinking. It supports asynchronous tools only. The client must keep listening for work after an audio turn ends and use the reported interaction status to distinguish ongoing processing from an idle session.
For example, reading the next available appointment from a simple test calendar may not need the same reasoning as resolving several scheduling constraints. Run the same task with the same test data before choosing. Record whether extra reasoning improves correct outcomes enough to justify any additional delay or implementation work.
A model's published benchmark position is not a measurement of your workflow. Your caller's vocabulary, background noise, tool responses and handoff process can change what happens in practice. Treat the pilot as a comparison against your current process, with a named reviewer and an explicit definition of success.
Estimate the cost of an actual conversation
The current Gemini API pricing table lists audio input at US$0.005 per minute and audio output at US$0.018 per minute for both models. These are component rates, not an all-in price for operating a telephone assistant.
Hypothetical calculation: if a test uses six billable input-audio minutes and four billable output-audio minutes, those components total US$0.102: six multiplied by $0.005, plus four multiplied by $0.018. This illustration assumes the stated billable quantities; it does not predict how a real ten-minute session will be billed.
Add any relevant text or visual input, other tool usage, telephony, hosting, monitoring, retries and human review. Check actual billing against the pilot logs. A cheaper conversation that needs lengthy correction may cost more to operate than a more reliable alternative.
Also verify the data terms for the exact service and tier. Google's pricing page distinguishes how free and paid API content is used to improve its products. That distinction does not answer every retention, access or business-data question, so review the applicable terms before introducing real customer information.
Update the integration, not just the model name
The standard model's migration notes identify configuration changes from the earlier Flash Live preview. They include removing unsupported thinking settings, changes to function-calling behaviour, permanently enabled proactive audio and removal of the earlier affective-dialogue setting. Ask your developer to review those notes against the existing client before changing production traffic.
For Extended Thinking, an especially important detail is that an ended turn does not necessarily mean all work has finished. A booking application should track the request and its saved result independently of the spoken response. It should also know what to do if the user changes the date while an earlier lookup is still running.
The Live API overview describes a stateful connection, meaning the application maintains an ongoing session. For browser clients, Google recommends short-lived ephemeral tokens rather than exposing a standard API key. Have the implementer check session handling, permissions and recovery as part of the pilot.
If your current need is simply to turn recordings into text, keep that evaluation separate. Our Grok Voice Transcribe 2.0 business guide covers a transcription-oriented decision. Likewise, conversation compaction addresses how context is managed, rather than whether a voice workflow completed the requested action.
A worked pilot: a repair shop appointment request
Imagine a fictional repair shop testing an internal assistant before customers can use it. The test calendar contains invented appointments, and the assistant can propose a slot but cannot charge a card or send a real customer message. A staff member approves the final test booking.
The caller asks for Thursday morning, then interrupts to say Friday afternoon. The application must avoid treating an unfinished Thursday lookup as the final instruction. After the caller agrees to an available Friday slot, the reviewer checks the test calendar for exactly one correct booking.
That is the basic success condition. Now make the scenario less tidy: return an unavailable slot, delay a tool response, disconnect the session, then repeat the request. Check whether the application explains what is pending and avoids creating a second booking. These are proposed tests, not claims about observed Gemini behaviour.
Use this compact prompt to prepare fictional test cases:
Create six fictional appointment-assistant test cases.
Include an interruption, a changed date, an unavailable slot,
a delayed tool response, a disconnect and a repeated request.
For each: caller request, expected saved result, unacceptable
action and evidence the reviewer must check.
Use invented details. Do not claim any test has passed.A prompt is only a planning aid. The application needs actual controls for permissions, approval and duplicate prevention. Give the developer a test calendar and a request identifier to trace each attempt; have a reviewer inspect the resulting record. Keep the original request, any correction and the final outcome together so a reassuring transcript cannot conceal the wrong booking.
Decide what would make the pilot worth keeping
Set the decision rules before the trial. Use separate observations for conversation quality and task completion, so an engaging voice does not dominate the review.
- Understanding: did the assistant preserve the latest request, names and other required details, and ask when something was unclear?
- Outcome: did the connected test system contain exactly the intended result, with no duplicate or unauthorised action?
- Recovery: could a person understand the state after an interruption, timeout or tool failure and take over?
- Effort and cost: how much review and correction was needed, and what did the complete attempt cost?
Include the languages, accents and noisy environments your actual users are likely to bring. Do not extrapolate from one clean recording. Choose a small test set you can inspect thoroughly, record the conditions and repeat failed cases after changes.
For this hypothetical appointment pilot, a sensible proposed release rule is no duplicate bookings or unapproved writes in the reviewed cases, a clear handoff for unresolved requests, and a demonstrable improvement over the current process. Passing that small test set would justify a controlled next step, not a claim of universal reliability.
If the pilot mainly adds monitoring work or leaves reviewers guessing whether actions finished, keep the existing process while you improve the implementation. The release creates another option to evaluate; it does not make a voice agent necessary for every business.
Reader Q&A
What is Gemini 3.8 Live?
Gemini 3.8 Live is Google's September 2026 voice model for real-time conversations and connected tasks. Businesses can evaluate it through the Live API, while availability in Google's own products follows separate rollout and account rules.
How does Extended Thinking differ?
Extended Thinking is designed for more involved reasoning during a voice conversation. Its integration must handle background work and asynchronous tool calls. Choose it only if a controlled trial shows a useful improvement for your particular task.
Does a spoken confirmation prove an action finished?
No. Check the connected system's saved result. A reassuring reply, a completed audio turn and a successfully recorded booking are different events. Your application should handle pending work, failures and repeated requests explicitly.
What does the audio API cost?
On 22 September 2026, Google's pricing table lists US$0.005 per minute of audio input and US$0.018 per minute of audio output for these models. Text, visual input, other services and operational work can add to the total.
Can I replace a transcription tool with a voice agent?
Only if your task actually needs a two-way conversation. Producing a transcript and carrying out an interactive workflow are different requirements. Compare transcript accuracy, speaker handling and export needs separately before changing a transcription process.
What is a sensible first business test?
Start with a reversible internal task using fictional data and a test system. Include an interruption, a correction, an unavailable option, a tool failure and a repeated request. Review the saved outcome and human effort before considering real customers.
The next step
Choose one low-risk conversation, one test system and one reviewer. Confirm access, run the awkward cases and inspect the saved outcomes before expanding the scope. That gives the Gemini 3.8 Live announcement a useful business context: evidence about a particular job, rather than confidence borrowed from a smooth demo.
