Google launches Gemini 3.8 Live models with background reasoning for voice agents

Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to enable more intelligent near real-time voice agents. [1][3] The 3.8 Live model combines conversational intelligence with visual grounding, while the Extended Thinking model is designed for high-complexity tasks requiring multi-step reasoning. [1][3]

Developers can use the models via the Live API to build agents capable of executing API and tool calls in the background while continuing to stream audio to the user. [2] The Extended Thinking model supports configurable thinking to handle reasoning in the background while responding or narrating progress. [2]

Gemini 3.8 Live Extended Thinking captured the number one spot on the Artificial Analysis Speech to Speech Quality Index with a score of 82.6. [1][3] It also led in agentic task completion with 68.6% on the τ-Voice benchmark and 35.1% on Sierra’s τ-Voice-banking benchmark. [1][3]

The models are available via the Live API, with pricing at $0.005 per minute for audio input and $0.018 per minute for audio output. [2] The release expands Google's developer suite for building real-time, voice-first product experiences. [2]

The introduction of asynchronous function calling in Gemini 3.8 Live addresses a key friction point in voice AI by allowing agents to fetch data without pausing the conversation. The Extended Thinking model's performance on the Artificial Analysis Speech-to-Speech Quality Index positions it strongly against other frontier voice models. What the sources don't address: how the latency of the extended reasoning process affects the user experience in practical voice applications.