Google Launches Gemini 3.8 Live and Live Extended Thinking for Real-Time Voice

It is a speech-to-speech model from Google that converts audio directly to audio without a text intermediary, reducing latency and allowing natural conversation with mid-speech interruption.

Both are bidirectional voice models. Gemini 3.8 Live is exposed through a WebSocket endpoint in the Google AI API with an Extended Thinking variant for deeper reasoning, while GPT-Live runs on OpenAI interfaces. Practical differences show up in Arabic dialect quality, pricing, and Google Cloud integration.

It suits early pilots and multilingual contact centers, but test multiple Arabic dialects before production and review data governance and local compliance requirements per country.

It gives the model more reasoning time before responding, useful for complex multi-step questions, though it may increase response latency versus the real-time mode.