Google DeepMind has unveiled Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two new additions to its flagship model family that push the boundaries of real-time multimodal AI interaction and reasoning depth.
Gemini 3.8 Live is designed for real-time, streaming interactions — think live voice conversations, real-time video understanding, and dynamic tool use. It processes audio, video, and text simultaneously, enabling use cases from live meeting transcription with contextual follow-ups to real-time visual analysis during video calls. The model maintains low latency even during complex multi-turn exchanges, making it suitable for production applications that demand responsiveness.
The Extended Thinking variant adds a deeper reasoning layer on top of the Live architecture. Before responding, the model can engage in an internal chain-of-thought process, weighing multiple approaches and refining its answer iteratively. This makes it particularly suited for complex problem-solving, coding tasks, and analytical work where accuracy matters more than raw speed.
DeepMind says the new models achieve significant improvements in benchmark performance across reasoning, coding, and multimodal understanding tasks. The Extended Thinking variant in particular shows substantial gains on math and science benchmarks where step-by-step reasoning is critical to getting the right answer.
The models are available through Google AI Studio and the Gemini API, with pricing structured around input and output token counts. DeepMind has also published detailed model cards documenting capabilities, limitations, and safety evaluations for each variant.