Vercel AI Gateway adds Google's Gemini 3.8 Live real-time voice models

Google's Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two models built for real-time spoken interaction, are now available through Vercel's AI Gateway. According to a changelog post published on September 15, both models are aimed at voice assistants, conversational products, and any application that responds in audio.

What the models do

The base model, exposed as google/gemini-3.8-live, handles real-time audio and adds visual grounding, Vercel says. It switches automatically across 97 languages and can run tool calls in the background while a conversation keeps flowing, so a data lookup does not stall the dialogue.

The Extended Thinking variant, google/gemini-3.8-live-extended-thinking, layers multi-step reasoning on top of that. Vercel describes the reasoning as running in parallel with speech: the model can acknowledge a request and narrate its progress instead of going silent while it works. For voice products, that is the difference between audible feedback and dead air.

How to connect

Both models are reached through the AI SDK's realtime API. The setup Vercel documents has three moving parts: install the Gateway provider package (@ai-sdk/gateway) alongside a WebSocket client such as ws, mint a short-lived token for the chosen model, and open a WebSocket using the connection configuration the model adapter returns. That adapter also serializes outgoing client events and parses incoming server events, so developers work with structured events rather than raw frames.

Session configuration is sent as a session-update event, where you can request audio output and audio transcription. On the Extended Thinking model, reasoning depth is configured through provider options, and Vercel notes that exactly one of thinkingLevel or thinkingBudget must be supplied — not both.

On the response side, the stream includes events such as audio-transcript-delta for incremental transcript text, plus response-done and error events that signal when to close the socket or surface a problem. Vercel points developers to its realtime quickstart for the full event reference, and both models can also be tried in the model playground without writing any code.

What the Gateway adds

AI Gateway is Vercel's unified layer for calling models from multiple providers. Beyond routing, it handles usage and cost tracking, retries, failover, and performance optimizations, which Vercel says delivers higher uptime than calling a provider directly.

Why it matters

Real-time voice is one of the most demanding deployment targets for language models. Users expect fast, interruptible turns and a natural feel when a model needs to think or fetch data mid-conversation. The capabilities Vercel lists — parallel reasoning, background tool calls, automatic multilingual switching — map directly onto those expectations, and the Extended Thinking variant in particular targets the awkward silences that break voice experiences on other models.

The packaging matters as much as the models. Teams already building on the AI SDK can add a speech mode without adopting a second vendor SDK, a separate billing relationship, or their own reliability plumbing, since failover and usage tracking arrive with the Gateway. For developers of voice assistants and audio-first interfaces, the integration work shrinks to changing a model identifier and handling streaming events.