Google / DeepMind Gemini: Live API, Deep Research Agent and Translate upgrades

Google / DeepMind announced multiple Gemini updates focused on native audio, real-time translation and research agents. - Realtime speech-to-speech translation added to Google Translate. - Translate supports over 70 languages in beta on Android. - Translate modes: Continuous Listening and Two-way Conversation. - Initial region support: US, Mexico, India; iOS and more countries in 2026. - Google launched Gemini Live API based on Gemini 2.5 Flash Native Audio. - Gemini Live API provides native audio processing and sub-step latency reduction. - API supports WebSocket single-connection media streams. - Five humanizing capabilities: emotion detection, smart interruption, adaptive tone, tool calling, persistent memory. - Quickstart templates: Vanilla JS and React. - Use cases demoed: real-time business advisor, multimodal customer service, game voice assistant. - DeepMind released Gemini Deep Research Agent for developer research tasks. - Deep Research Agent autonomous planning, multi-source integration and citation support. - Agent benchmark scores: DeepSearchQA 66.1%; Humanity's Last Exam 46.4%; BrowseComp 59.2%. - Interactions API is a unified interface for models and agents and is available in Google AI Studio.
Key insights
Gemini powers real-time speech-to-speech translation in Translate
Google announced real-time speech-to-speech translation integrated into Google Translate, supporting Continuous Listening and Two-way Conversation modes and over 70 languages in beta on Android. Initial region availability is United States, Mexico and India, with iOS and additional countries expected in 2026. The feature enables live translated audio through any headphones and automatic language detection for both interlocutors.
Gemini Live API introduces native audio processing and low-latency multimodal streams
Google launched the Gemini Live API powered by Gemini 2.5 Flash Native Audio, enabling native end-to-end audio processing (listen and generate audio) without STT/LLM/TTS stitching. The API uses a single WebSocket connection for media streams, reduces latency, supports video and visual context, and provides templates (Vanilla JS, React) and demos for rapid developer adoption.
Gemini Deep Research Agent aimed at autonomous, verifiable developer research
DeepMind released Gemini Deep Research Agent built on Gemini 3 Pro, trained for multi-step research workflows. The agent autonomously plans searches, integrates private documents with web results, and produces reports with citations. It reported SOTA results on internal benchmarks including DeepSearchQA (66.1%), Humanity's Last Exam (46.4%) and BrowseComp (59.2%). The agent is accessible via the Interactions API and Google AI Studio.
Developer tooling includes unified Interactions API and demos
Google provided a unified Interactions API to combine models and agents in one endpoint, plus Quickstart templates and three representative demos (real-time business advisor, multimodal customer support, and game voice assistant). The API emphasizes controllability, JSON outputs and enterprise stability for production usage.
Gemini's humanizing audio features aim to improve conversational realism
Google highlighted five capabilities intended to make voice agents feel more human: emotional tone detection and adaptation, intelligent interruption and turn-taking, tool calls during voice, persistent memory across interactions, and enterprise-grade reliability for production use.