On September 15, Google put two new real-time voice models on the table at once — Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking — billed as "the most conversational Gemini yet", rolling out today across the Gemini API, Google AI Studio, Gemini Enterprise, and Search Live.
What the numbers say: how the two variants split the work
Both models orbit a single goal: making conversation feel like talking to a person. The standard Gemini 3.8 Live emphasizes cost and scalability, taking second place on user preference in the Speech Agent Arena. The Extended Thinking variant pushes into complex tasks — it tops Artificial Analysis's Speech to Speech Quality Index at 82.6, hits 35.1% on Sierra's τ-Voice-banking agentic benchmark, and lands 97.7% on the Big Bench Audio reasoning benchmark. On ServiceNow's EVA-Bench, both versions jointly push the accuracy-vs-experience Pareto frontier for voice agents to a new position.
Capabilities: listen, think, speak, see, switch, background-tool — all in one shot
Google surfaced several concrete capabilities this round: real-time visual input, near-real-time voice response, and automatic detection and switching among 97 supported languages mid-conversation. The Extended Thinking variant's standout feature is "thinking while talking" — the model runs multi-step reasoning in the background while emitting language cues like "Let me check that..." to keep the conversation flowing, only resuming the full response when the background task completes. That means developers on the Gemini Live API can ship voice agents that do multi-step planning without ever freezing the live conversation. Tool calling was rewritten as well: both versions run tool calls and API requests in the background without interrupting the voice thread, with the front-end transitioning naturally on a single "let me check that for you". SynthID watermarking is baked into every audio output. Google also named eight ecosystem partners already shipping integrations on top of Gemini Live API — Agora, LiveKit, Pipecat, Vercel, LangChain, Fishjam, Vision Agents.
Pricing, industry meaning, and rollout cadence
Worth pulling apart: pricing. Google didn't publish raw numbers in the blog, but Artificial Analysis's cost-per-hour-of-input-audio axis and the Big Bench Audio value-positioning both place Extended Thinking in the "frontier but cheap" quadrant. From an industry angle, this release moves real-time voice out of the "TTS + ASR + LLM stitched together" era into an end-to-end unified-model era — Gemini 3.8 Live simultaneously handles listening, reasoning, speaking, visual reading, language switching, and background tool execution, collapsing what used to require several independent systems into a single set of weights. For teams building customer service, education, and companion voice products, the infrastructure assembly work gets compressed significantly; differentiation now has to live in product logic and user experience.
The model card went live on Google DeepMind's model-cards repo alongside the Safety & Responsibility report and the SynthID policy. Gemini Enterprise's private preview also opened today — making this the first voice model enterprises can drop into production without bolting on NLU/ASR/TTS separately. Salesforce, Genspark, and Lumeris, already shipping Gemini 3.8 Live integrations, will be the testbed for whether these capabilities survive real scenarios.