Google Releases Gemini 3.8 Live and 3.8 Live Extended Thinking for Production Voice Agents

The two models divide the work by task complexity. Gemini 3.8 Live is designed for scale and cost efficiency, combining conversational intelligence with fluid dialogue and visual grounding. Gemini 3.8 Live Extended Thinking targets high-complexity work, adding multi-step reasoning that runs while the model is speaking. Google presents both as a streamlined alternative to cascaded speech pipelines that chain automatic speech recognition, a language model and text-to-speech, a configuration in which reasoning and tool execution often interrupt the conversation.

On benchmarks reported at launch, Gemini 3.8 Live Extended Thinking took first place overall on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6. It led agentic task completion with 68.6 percent on τ-Voice and 35.1 percent on Sierra's τ-Voice-banking benchmark, and scored 97.7 percent on Big Bench Audio, a reasoning benchmark for audio models. Gemini 3.8 Live placed second in the Speech Agent Arena, a human preference evaluation. On ServiceNow's EVA-Bench, Google said the models push the Pareto frontier for complex workflows, balancing task accuracy against conversational quality as measured through the Live API on the Gemini Enterprise Agent Platform.

The Live API exposes five capabilities. Asynchronous function calling lets the model execute API and tool calls in the background while audio responses keep streaming to the user. Visual context allows near real-time processing of live visual inputs, so an agent can respond to what a user both says and sees. Alphanumeric precision handles confirmation codes, claim numbers and technical data, a common failure point in voice systems. Multilingual support detects and switches between 97 languages mid-conversation while maintaining accent consistency. Incremental content updates merge real-time audio with structured data to return context-aware responses.

Extended Thinking adds configurable thinking for multi-step reasoning in the background. It reasons and speaks at the same time, using early verbal cues such as "Let me check that" to acknowledge a prompt, then narrates progress step by step while long-running tasks execute. Google's demonstrations showed the model converting sketches plus voice feedback into working React components and coordinating multi-step bookings.

Both models are priced at $0.005 per minute for audio input and $0.018 per minute for audio output, an estimate Google bases on $3 per million input tokens and $12 per million output tokens. Developers can also build through Live API integration partners handling real-time media streaming infrastructure, including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents. Google is additionally partnering with Salesforce, Genspark and Lumeris, which cited the models' latency, fluidity and tool calling. Example applications are available on GitHub, and all generated audio carries Google DeepMind's imperceptible SynthID watermark.