AIBRIEF
The new 'Extended Thinking' mode lets Gemini Live keep talking while it reasons or runs tools, and the assistant now auto detects which of 97 languages a user is speaking.
Google's Gemini Live voice assistant can now keep talking while it does the harder work behind the scenes. The new Extended Thinking mode lets the model reason through a multi-step request and run tools in the background rather than going silent until it has a finished answer. When it needs a moment, it says things like "Let me check" instead of freezing (Google blog).
The base Gemini 3.8 Live update adds two reader-facing details: the assistant now auto-detects which of 97 supported languages a user is speaking and switches mid-conversation without a manual toggle, and it processes photos and camera input in near real time so a user can point at something and ask about it.
The change Google is selling is that voice stops being a one-shot query tool and behaves more like a person who talks while they work. That is the experience a user gets the next time they open Gemini Live to plan a trip, work through a recipe from a fridge photo, or summarize a long document.
Google says Extended Thinking ranks first on Artificial Analysis' Speech to Speech Quality Index at 82.6, hits 97.7% on Big Bench Audio, and 68.6% on Sierra's τ-Voice. Those figures are Google-reported, not independently re-run. On ServiceNow's EVA-Bench voice-agent test, the company says the models "push the Pareto frontier," a claim qualified even in Google's own framing.
What remains open is whether the verbal bridges hold up on messy audio. Hacker News testers called Gemini Live strong on translation and natural-sounding speech but weak on coding, and several asked whether Google is still shipping a "Pro" frontier model (HN thread). Pricing, latency, and regional availability across the Gemini app, Workspace, Search, the Live API, and Gemini Enterprise Agent Platform are not in the post.