Gemini 3.8 Live and Extended Thinking bring background reasoning to voice agents
Google's two new live dialogue models pair spoken conversation with background tool execution. Developer rollout and enterprise preview access differ.
AI Neural Narration
48kHz StudioFish Audio Neural Engine · Natural editorial narration
Key Takeaways
- check_circleThe Live family includes a standard model and an Extended Thinking variant.
- check_circleBackground work needs a clear completion signal in any voice interface.
- check_circleCheck access in the intended account before planning a rollout.
What Google announced
Google announced the two models on 15 September 2026. Live targets efficient dialogue; Extended Thinking handles deeper reasoning. Google describes background tool execution, visual grounding and switching between 97 languages.
Developer rollout is through the Gemini API and AI Studio. Gemini Enterprise access is a private preview. Consumer availability varies across Search Live, Gemini Live and Workspace subscriptions.
A useful conversation still needs reliable actions
Our assessment: speaking while a task runs can make an assistant easier to use, but it creates a design problem. A friendly acknowledgement must remain distinct from confirmation that an action succeeded. For a booking assistant, the customer should hear that a reservation is confirmed only after the booking system returns a successful result.
Evaluate each variant against a small set of real conversations. Include interruptions, corrections, unfamiliar place names and requests that need several tool calls. Record how often the assistant completes the task, how long the customer waits, and whether the spoken answer agrees with the system's final state. A smooth demo does not answer those questions.
What to check before changing an existing voice service
Keep your current agent as the comparison point. Give both systems the same tasks and the same permitted actions. Measure the cost of a completed conversation, including retries and human handoffs, rather than comparing a token rate alone. For complex requests, decide whether the additional reasoning actually saves staff time.
Prepare a clear fallback when a connection drops or a tool fails. The interface should tell the user what remains unfinished and make it easy to reach a person. For multilingual services, ask fluent speakers to review representative calls in the languages your customers use. Treat the launch announcement as a starting point for that evaluation.
Frequently Asked Questions
Are both Live models publicly accessible to developers?
Google lists rollout through the Gemini API and Google AI Studio. Enterprise preview and consumer subscription access have separate conditions.
Which variant should a business evaluate first?
Start with the tasks your agent handles most often. Compare the standard and Extended Thinking variants on completion rate, waiting time and cost using the same conversations.