Google introduced Gemini 3.8 Live and Live Extended Thinking on September 15, adding voice models that can keep a conversation going while handling background tasks. The Extended Thinking version combines spoken replies with ongoing reasoning, allowing a user to hear progress while a more complex request is being worked through.
Both models are based on Gemini 3 Pro and take audio, images, video and text as inputs, producing speech and text. A spoken request can require more than an answer generated from the conversation: completing a booking, for example, involves checking information and making calls to external software. Those steps create waiting time between a request and its completion.
Google says the standard Live model can run external tools while continuing to talk. Extended Thinking adds simultaneous reasoning and speech for more demanding sequences: it first acknowledges the request, then provides verbal updates as background work proceeds. This describes the interaction a user receives. The launch material does not explain enough of the internal scheduling to establish how reasoning and speech share computing resources, or whether a progress statement accurately reflects each completed operation.
To test completion, Google's evaluation report includes Sierra's tau-cubed Banking benchmark, which asks agents to search a large, unstructured knowledge base and execute several software actions for banking workflows. Gemini models were tested through the Gemini application programming interface (API), the connection used by developers to send requests, with high thinking effort and default sampling, using one attempt per task. Extended Thinking scored 35.1%, compared with 11.3% for Gemini 3.1 Flash Live at high thinking effort. That is a substantial improvement within the reported test, though most tasks still did not succeed.
Conversational experience also needs a separate measurement. In ServiceNow's EVA-Bench, automated agents conduct complete spoken conversations, with task accuracy and user experience scored separately. Google's published plot places standard 3.8 Live higher on experience and Extended Thinking at high effort higher on accuracy. The comparison shows why a model choice depends on the workflow. Google's model card also lists possible fabricated answers, occasional slow responses and timeouts; these remain relevant even when the conversation sounds natural.
Developers can begin evaluating that choice through the Gemini API and Google AI Studio, where Google says both models started rolling out on September 15. Enterprise access is in private preview, with the customer-experience offering announced as forthcoming. Availability therefore varies by product channel.
For teams building voice services, the release provides an option for keeping callers informed during a sequence of software actions. Deployment evidence should now connect those spoken updates to verified task completion, including how often the service stalls or needs a person to take over.