Google Launches Gemini 3.8 Live With Extended Thinking for Voice Agents

Voice agents keep promising “just talk” workflows—then freeze while the model thinks. On September 15, 2026, Google’s Gemini blog introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two live-dialogue models built to keep conversations moving while harder work runs in the background.

What Google claims on the leaderboards

Google says Extended Thinking takes the #1 overall spot on Artificial Analysis’ Speech to Speech Quality Index (82.6), with 68.6% on τ-Voice and 35.1% on Sierra’s τ-Voice-banking, plus 97.7% on Big Bench Audio. Gemini 3.8 Live is pitched as the scale/cost-efficient sibling and ranks second in Speech Agent Arena preference in Google’s write-up. Treat those as vendor-cited benchmarks until third-party replications land.

Live features that matter for agents

Both models emphasize parallel reasoning, near-real-time visual grounding, automatic switching across 97 languages mid-conversation, and background tool/API calls while the assistant keeps chatting. Extended Thinking is the deep-work variant: early verbal acknowledgements (“Let me check that…”) plus live progress narration during multi-step tasks.

Where it ships

Rollout started September 15: Gemini API and Google AI Studio for developers; private preview in Gemini Enterprise; Search Live for 3.8 Live; Gemini Live plus Workspace Docs/Gmail/Keep Live surfaces for eligible AI Pro/Ultra subscribers. Audio outputs carry SynthID watermarking.

Rocks take

Flag this as a first-party product launch. The useful buying question for builders is latency + tool reliability under interruption—not the Speech-to-Speech headline alone. Source: Google Gemini blog, September 15, 2026.

Why “talk while tools run” is the product bet

Prior live models often forced stop-and-wait whenever a calendar booking or spreadsheet lookup kicked off. Google’s framing is that 3.8 Live keeps the social contract of conversation while Extended Thinking handles multi-step agent work. That only wins if background failures are narrated honestly instead of silently stalling.

Partners named in the post (Agora, LiveKit, LangChain, Salesforce, and others) matter more for production voice stacks than a single keynote demo.

What to verify before you rebuild a voice agent

Ask for interruption handling under noisy audio, language-switch false positives, and how SynthID interacts with your compliance logging. If you are on Workspace Live, confirm which SKUs actually get Extended Thinking versus base Live.

Enterprise private preview timing can lag consumer Gemini Live—budget a canary, not a fleet cutover.

FAQ

Is this only in the Gemini app? No—API/AI Studio, Search Live, Enterprise preview, and select Workspace Live surfaces are called out.

Are the benchmarks independent? Index names are third-party; the scores in the blog are Google-reported.