Google's Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are now available on Vercel's AI Gateway. The base model handles real-time audio, visual grounding, automatic language switching across 97 languages, and background tool calls mid-conversation. The Extended Thinking variant runs multi-step reasoning in parallel with speech, letting it acknowledge requests and narrate progress without pausing the conversation.
Access both through the AI SDK's realtime API using the Gateway provider and a WebSocket client. The implementation pattern: mint a short-lived token, open the WebSocket, then use the model adapter to serialize and parse realtime events. The distinction between the two models, specifically how Extended Thinking manages parallel inference without blocking audio output, is worth reading in the original for the technical specifics.
AI Gateway wraps both models with unified usage tracking, cost monitoring, retries, and failover, targeting uptime above what Google's API provides directly. The model identifiers are google/gemini-3.8-live and google/gemini-3.8-live-extended-thinking. Both are live in the model playground now.
[READ ORIGINAL →]