Sign InOpen Brain
VercelEngineering PostOfficial Source

Gemini 3.8 Live models now available on AI Gateway

Two Gemini 3.8 Live models add real-time audio to AI Gateway; the extended variant can reason alongside speech while the base model keeps tool calls in the background.

Vercel · Sep 15, 2026
Open Source Open MarkdownOpen JSON
Source Summary

AI Gateway now exposes **Gemini 3.8 Live** and **Gemini 3.8 Live Extended Thinking** through the AI SDK's realtime API. Both handle spoken interaction; the base model adds visual grounding, background tool calls, and switching across **97 languages**.

Practical Implication

For voice agents, use short-lived tokens and the supplied WebSocket adapter, then choose the extended model only when multi-step reasoning must continue alongside speech. Gateway can centralize usage, cost, retries, and failover.

Agent-Ready Context
AI Gateway now exposes **Gemini 3.8 Live** and **Gemini 3.8 Live Extended Thinking** through the AI SDK's realtime API. Both handle spoken interaction; the base model adds visual grounding, background tool calls, and switching across **97 languages**.

For voice agents, use short-lived tokens and the supplied WebSocket adapter, then choose the extended model only when multi-step reasoning must continue alongside speech. Gateway can centralize usage, cost, retries, and failover.

The material gives no latency, pricing, or quality measurements, and realtime support is exposed through an **experimental API**. Extended Thinking also requires choosing exactly one thinking control: level or budget.
Connected Context · Feed7 Judgment

This broadens realtime voice routing with visual grounding, background tools, multilingual switching, and an explicit choice between ordinary and extended reasoning. Against existing live models, it makes concurrent speech-and-work capability a selectable architecture rather than a unique feature. Missing latency, cost, and quality measurements—and the experimental API—leave model choice dependent on matched voice workloads.

Context Map
modelaudio#reasoning#tool-use
Uncertainty
The material gives no latency, pricing, or quality measurements, and realtime support is exposed through an **experimental API**. Extended Thinking also requires choosing exactly one thinking control: level or budget.