Google Ships Gemini 3.8 Live Voice Models — Second Gemini 3.8 Release in Two Weeks
On September 15, 2026, Google launched Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its newest speech-to-speech voice models, available day one in the Gemini API, Google AI Studio, Search Live, and the Gemini app (Google blog). This is the second Gemini 3.8 release inside two weeks — Gemini 3.8 Flash and 3.8 Flash Cyber shipped September 2 — part of a cadence where Google is no longer shipping one flagship a year.
What actually shipped:
- "Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are our most advanced live dialogue models yet. Major upgrades in intelligence and parallel reasoning make them more intuitive to collaborate with and use to execute complex tasks using your voice."
- Benchmark claims: Gemini 3.8 Live Extended Thinking takes "#1 overall spot on Artificial Analysis' Speech to Speech Quality Index (82.6)", leads agentic voice tasks at "68.6% on τ-Voice and 35.1% on Sierra's τ-Voice-banking benchmark", and scores "97.7% on Big Bench Audio."
- Real-time behavior: processes visual inputs near real-time, auto-detects and transitions across "97 supported languages mid-conversation", and executes tools/API calls in the background while conversation continues. Extended Thinking "reasons and speaks simultaneously," using early verbal cues like "Let me check that…" and live progress narration.
- All output audio carries SynthID watermarking; a model card is published.
Why it matters: Voice is emerging as the next interface battleground — Google is going directly at the live-agent layer (Search Live means ordinary users hit the new model today, not in a beta window) while OpenAI and Apple make their own voice pushes (see Apple Ships Siri AI Beta on Next-Gen Apple Intelligence — Gemini-Built Models, EU and China Left Out). Notably, Google is shipping at this pace during the week its CEO endorsed Amodei's call to slow the frontier (see The Pacing Debate Ships Product: OpenAI Kills GPT-6.1, Nvidia Ships Agent Containment, the Market Starts Pricing the Slowdown) — the pacing talk has so far changed no one's shipping cadence. The claims are anchored to third-party leaderboards (Artificial Analysis, Sierra, ServiceNow's EVA-Bench) rather than self-reported evals, which makes them more falsifiable than Meta's Muse Spark parity claims (see Meta Launches Muse Consumer Agent on Muse Spark 1.3 — Subscription Economics and a Trust Test).