Google Launches Gemini 3.8 Models That Can Reason While Talking

Google has launched Gemini 3.8 Live models that can reason while they talk. The Extended Thinking model, part of this launch, allows voice assistants to continue conversing with users even when complex tasks are being performed in the background.

The new models, released on September 15th, bring real-time voice interactions and deeper background reasoning capabilities to Google's latest generation of AI agents. Gemini 3.8 Live Extended Thinking changes a fundamental assumption in voice-agent design: that a spoken response indicates the end of a task is no longer true.

Extended Thinking requires asynchronous function calls, which means tools can continue working while audio streams and responds. This architecture fits with Google's broader push toward agents that work across software and services, connecting to apps, files, and MCP servers.