Google ships Gemini 3.8 Live and 3.8 Live Extended Thinking

What happened

Google’s newsroom post, published September 15, 2026, describes two new additions to its Gemini Live lineup. Gemini 3.8 Live is pitched at scale and cost efficiency, combining conversational fluency with visual grounding so the model can reference what a camera or screen share shows during a spoken exchange. Gemini 3.8 Live Extended Thinking targets harder tasks: it reasons and speaks in the same turn, using verbal filler such as “let me check that” while it works, and narrates progress through multi-step jobs like coordinating a booking or turning a sketch into working code.

Both models can detect and switch between many supported languages without the user restarting the conversation, and both can hand off tool calls or API requests to run in the background while the spoken exchange continues uninterrupted. Google frames this as the difference between a voice assistant that pauses while it works and one that keeps talking.

Availability is staggered by audience. Developers can access both models today through the Gemini API and Google AI Studio. Enterprise customers get Gemini 3.8 Live in private preview through Gemini Enterprise, with Extended Thinking following for Gemini Enterprise for Customer Experience and Workspace business customers. Consumers reach the plain Live model through Search Live, and Extended Thinking through the Gemini app and, for AI Pro and Ultra subscribers, Docs, Gmail, and Keep.

The numbers

Google’s post cites third-party benchmark placements rather than internal figures. It says Extended Thinking took the top spot on Artificial Analysis’ Speech to Speech Quality Index and posted a strong score on Big Bench Audio, while the plain Live model placed second in the Speech Agent Arena ranking. Google did not publish new per-token API pricing alongside the announcement.

Why it matters

The two-tier split forces a concrete build choice on anyone integrating voice AI. A customer-facing assistant that mostly confirms requests and delegates to background tools has little reason to pay for Extended Thinking’s added reasoning step; a workflow that needs the model to talk a user through a multi-stage process, such as rebooking a flight across several conditional steps, is the case the slower model is built for. Because Google has not listed public per-token prices for either model, teams migrating from an earlier Gemini Live version cannot yet do a clean cost comparison and will need to test against their own traffic before switching.