Google launches Gemini 3.8 Live models optimized for real-time voice agents

Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to enable voice agents more effectively. [1] The company describes them as its "most advanced live dialogue models yet." [1]

Collectively referred to as Gemini 3.8 Audio, these models are optimized for latency-sensitive, high-volume tasks like real-time dialogue. [2] Built on the Gemini 3 Pro architecture, the models accept audio, images, video, and text inputs with a context window of up to 128K tokens. [2] The models output audio and text with a token limit of up to 64K. [2]

The introduction of the Gemini 3.8 Live models underscores a shift toward native audio capabilities optimized for real-time responsiveness. [2] By offering a 128K token input window and outputting up to 64K tokens, the architecture supports extended, multimodal interactions. [2] What the sources don't address: How the latency and cost of these new models compare specifically to previous iterations in real-world deployment.