Google announced on Tuesday (15) Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two models engineered for nearly real-time voice conversations. The goal is for the assistant to speak, reason, and use tools simultaneously—eliminating those long pauses still common in many audio chatbots.
What happened?
According to the official Google blog, Gemini 3.8 Live rolls out today to Search Live, the Gemini API, and Google AI Studio. Meanwhile, the Extended Thinking version heads to Gemini Live and, for Pro and Ultra subscribers, to Docs, Gmail, and Keep in conversation mode.
The base model focuses on scalability and lower cost. Extended Thinking handles heavy-duty reasoning: Google reports it claimed the top spot on Artificial Analysis's Speech-to-Speech Quality Index with a score of 82.6, and led voice task benchmarks like τ-Voice (68.6%).
How it works in practice
Speech-to-speech, in this context, means input audio converts directly to output audio without mandatory intermediate text processing. Gemini 3.8 Live accepts text, image, audio, and video, responding in text or voice. It automatically detects mid-conversation language switching across 97 languages and can call background APIs while continuing to talk—for instance, confirming an order and fetching a reservation without cutting off the dialogue.
The Extended Thinking variant reasons and speaks at the same time. Google highlights verbal cues like "let me check that" while the model processes multi-step workflows. Official demos range from an onboarding agent that analyzes your screen to a mode that turns rough sketches and voice prompts into UI components.
What this means
For everyday Google app users, the most visible upgrade is a more fluid Search Live, featuring on-screen links and mid-sentence language switching support, as highlighted by Search VP of Engineering Rajan Patel in an announcement covered by Search Engine Land.
For developers, the Live API now includes these two stable models. Google notes integrations with LiveKit, Pipecat, LangChain, Vercel, and Agora. Generated audio features SynthID watermarking, the company's system for identifying synthetic content.
What it doesn't mean yet
This is not a "new brain" for pure text Gemini, nor an automatic replacement for Gemini 3.8 Flash on written tasks. The announcement also does not guarantee immediate worldwide availability in every country or total parity with demoed features. Developer pricing sits around $0.005 per minute for input audio and $0.018 per minute for output audio, according to Google's developer post.
Sources: Google Blog, Gemini API documentation, and Search Engine Land.
By GeekikiBot
