Google Ships Gemini 3.8 Live Speech-to-Speech Models via Raw WebSocket API

Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, 2026 as speech-to-speech models reachable via a bare WebSocket endpoint, competing with OpenAI's GPT-Live family. On September 25, Google Cloud announced that Gemini 3.8 Live with Live Avatar, previously previewed at Google Cloud Next 2026, is now generally available in Gemini Enterprise, adding synchronized video avatars and background tool calling to the same voice stack.

• The models are accessed via wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1alpha.GenerativeService.BidiGenerateContent, usable with just the Web Audio API and no client library.
• Extended Thinking supports configurable reasoning levels and optional document context injected during an ongoing voice conversation.
• Gemini 3.8 Live with Live Avatar is now GA in Gemini Enterprise, generating lip-synced video avatars for web, mobile, and kiosk conversational agents; custom avatars remain allowlist-only.
• Live Avatar also runs tool calls and API requests in the background so the model can keep talking while backend tasks finish.
• Caveat: GA scope is limited to Gemini Enterprise and custom avatars still require allowlisting, so broader availability and independent performance testing remain unverified.

Bottom line Voice-agent builders now have a GA path to add lip-synced video avatars and background tool calling on top of the raw WebSocket speech-to-speech endpoint, though only inside Gemini Enterprise for now.