Gemini 3.8 Live is GA: voice-video agents, avatars for enterprise
A bidirectional voice+video API lands with a free tier and ~$0.005/min audio; the lip-synced 97-language avatar stays enterprise-gated for now.
Copy markdown
The Live API is the actual unlock
Gemini 3.8 Live went generally available on Google AI Studio and the Gemini API: a WebSocket stream that takes audio and video in and speaks back, with function calling on both gemini-3.8-live and an extended-thinking variant. Free-tier tokens on both models, then roughly $0.005/min audio in and $0.018/min out, so a 10-minute voice session runs about $0.23. That is enough to ship a real-time voice agent this weekend.
Avatars lip-sync in 97 languages, but they're gated
The headline feature, Live Avatar, adds an animated face that lip-syncs and switches languages mid-conversation across 97 of them, every frame watermarked with SynthID. The catch for solo builders: the avatar layer ships in Gemini Enterprise and custom avatars are allowlist-only, so today you get the voice-video API, not a bespoke face.
Extended thinking that reasons while it talks
The gemini-3.8-live-extended-thinking model reasons and speaks at the same time, so the agent keeps talking instead of going silent while it plans. Asynchronous tool calling runs background jobs without breaking the dialogue, useful for anything that has to look something up mid-sentence.
Elsewhere: Claude Code 2.1.282 fixes resumed-session drift
Anthropic shipped Claude Code 2.1.282 with a fix for resumed sessions silently re-sending earlier messages in changed form, plus a fix for sessions failing on web-search results the API cannot decrypt. A new maxProseWidth setting caps prose width in wide terminals.