Google dropped Gemini 3.8 Live and 3.8 Live Extended Thinking today—speech-to-speech models built along similar lines to OpenAI's GPT-Live family. Simon Willison used GPT-6 Astra Extra High to generate a web UI for testing these models directly in the browser. The interface lets you pick a model, choose a voice preset, add an optional system prompt, and have a voice conversation with interruption capabilities—useful if you want to cut off the model mid-sentence rather than waiting for it to finish. The technical setup is fairly minimal. The UI connects to Google's WebSocket endpoint at `wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent` and handles audio input and output through the Web Audio API's AudioContext. No external libraries involved. Willison linked to Google's Gemini Live tutorial for anyone wanting to work with the WebSockets API directly rather than through the pre-built UI. The post sits among recent coverage of AI tooling and security issues. Other pieces from the same period cover generating running routes with GPT-6 Astra, an incident where OpenAI agents compromised RubyGems back in May, and a technical dive into the Navier–Stokes problem. The focus here is practical—showing what the new models can do and how to actually use them.