Voice, Speech & Realtime AI

Google ships Gemini 3.8 Live to power enterprise voice agents that act, not just answer

Source: PYMNTS · Sep 22, 2026

Google has launched Gemini 3.8 Live, a voice-model family aimed squarely at enterprise voice agents. It comes in two tiers: a Live model tuned for scale and cost efficiency, meant to handle high volumes of real-time conversation cheaply, and an Extended Thinking variant priced for high-complexity work where deeper reasoning matters more than cost. Google frames both as production building blocks — not demos — for companies deploying voice agents that respond with near-real-time reasoning.

The significant shift is what these agents are being built to do. Voice AI is moving from answering questions to executing tasks — carrying out multi-step actions inside a live conversation rather than just retrieving an answer. That is a meaningfully harder bar, and it puts a premium on models that can reason quickly while a person is still talking.

For anyone working with conversational data, the relevance is direct. As voice agents get better and cheaper, more and more real human conversation flows through them, and the systems that train, evaluate and improve those agents are hungry for realistic, high-context dialogue. The better the agents get, the more valuable — and the more sensitive — the underlying conversational data becomes, which is exactly why consent and de-identification move from afterthought to prerequisite.

Key Points

• Gemini 3.8 Live (tuned for scale and cost efficiency) plus an Extended Thinking variant (for high-complexity, frontier-priced work), launched 2026-09-15
• Positioned as production building blocks for enterprise voice agents with near-real-time reasoning
• Signals voice AI moving from question-answering toward multi-step task execution
• Part of a broad September wave of realtime, conversation-native model releases