Wednesday, Sep 16 | --:--
Back to home

Google Ships Gemini 3.8 Live and Live Extended Thinking for Voice Agents

On September 15, 2026, Google launched Gemini 3.8 Live and 3.8 Live Extended Thinking, native speech-to-speech models for the Live API, Search Live, Gemini Live, and Workspace. Extended Thinking takes Artificial Analysis’ Speech-to-Speech Quality Index at 82.6.

Tech Insights Reporter 4 min read Mountain View, CA
Cover illustration for Google Ships Gemini 3.8 Live and Live Extended Thinking for Voice Agents

TLDR

Google on Tuesday, September 15, 2026 published Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking—native speech-to-speech models for production voice agents. 3.8 Live: scale and cost, visual grounding, background tool calls, auto-detect 97 languages. 3.8 Live Extended Thinking: high-complexity tasks; reasons and speaks at once. Independent: Extended Thinking #1 on Artificial Analysis Speech to Speech Quality Index (82.6); τ-Voice 68.6%; Sierra τ-Voice-banking 35.1%; Big Bench Audio 97.7%. Base 3.8 Live second on AA Speech Agent Arena. Dev: Gemini API + AI Studio Live API at $0.005/min audio in, $0.018/min out. Enterprise private preview in Gemini Enterprise. Consumer: 3.8 Live powers Search Live; Extended Thinking in Gemini Live and, for AI Pro / Ultra, Docs / Gmail / Keep. SynthID on all generated audio.

Two SKUs

SKU Job Where it ships today
3.8 Live Low-latency dialogue + visual + async tools API, AI Studio, Search Live, Gemini Enterprise (private preview)
3.8 Live Extended Thinking Multi-step reasoning while speaking API, AI Studio, Gemini Live, Docs/Gmail/Keep (Pro/Ultra)

Ranking: lab cites Artificial Analysis (82.6 #1 S2S Quality Index; Arena #2 for base Live) plus τ-Voice, Sierra banking, Big Bench Audio. EVA-Bench (ServiceNow): Google says the pair sit on the accuracy–quality Pareto. SiliconANGLE’s AA snapshot names GPT-Live-1-Astra and Grok Voice Think Fast 2.0 as the models it surpassed—do not treat that as an Arena Elo.

Product-line de-dupe: not Gemini 3.8 Flash / Flash Cyber (Sep 2), not 3.1 Flash Live (March), not 3.5 Transcribe (last month; 85+ languages, WER 4.0% streaming). Distinct from unfiled OpenAI GPT-Live-1 API (Sep 10 re-pass).

Why this story matters

Google is putting live speech on the same 3.8 family as Flash, and it is claiming the independent S2S board, not just a vendor table. Watch: whether Search Live actually sounds like Extended Thinking, and whether $0.005 / $0.018 holds once GPT-Live-1 is in the API.

Sources

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading