From Mexico: Google just launched Gemini 3.8 Live and 3.8 Live Extended Thinking — native speech-to-speech models with async tools, visual grounding, and (for Extended Thinking) background reasoning while the conversation keeps flowing. Rolling out today.
Google DeepMind shipped agentic video on Gemini Flash today. It hunts frames, audio, and transcript instead of fixed 1 FPS — vendor claims up to 88% fewer tokens.
Google DeepMind extended Gemini Co-Scientist into a lab partner that plans experiments, drives a CVD furnace, and drafts papers. Humans still load samples. A 450-review study cut severe fabricated results to 4%.
Google DeepMind piloted what it calls the first double-blind evaluation of a proprietary frontier model. Gemini Flash Lite and confidential benchmarks sat in one crypto enclave. Scores were not published.
Google shipped Gemini 3.5 Transcribe on 26 August: speech-to-text that cleans fillers, keeps self-corrections, and auto-detects 85+ languages. Here’s where it lives today, how to turn it on, and what it does to a ramble. Chrome talk-to-type is later, not now.