Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15 local time, two AI models built for spoken conversation. The first targets scale and cost efficiency, the second handles reasoning for complex work. Both started rolling out the same day through developer APIs and consumer-facing services.
Two models, two different jobs
What a voice agent needs depends heavily on the task. Handling large volumes of short exchanges, such as taking orders or answering routine questions, puts response speed and running cost front and center. Working through a multi-step procedure is a different problem, where a model that thinks carefully is worth a short wait.
Google split the lineup along exactly that line. Gemini 3.8 Live is designed for large-scale deployment and cost efficiency, pairing conversational fluency with visual grounding. Gemini 3.8 Live Extended Thinking targets high-complexity tasks and leans on multi-step reasoning.
Seeing while it talks, working while it talks
Gemini 3.8 Live stands out for how concretely it avoids breaking the flow of a conversation.
Language switching comes first. The model detects which of its 97 supported languages is being spoken and switches mid-conversation, so a speaker who mixes languages does not need to reset anything.
Vision comes next. Visual input is processed in near real time, letting the model answer based on what is actually in front of the user. Pointing a camera at a device and asking how to operate it is one of the intended cases.
Third is background execution of tools and API calls. Older voice assistants tended to fall silent while an external process ran. Gemini 3.8 Live acknowledges the request out loud first and keeps the conversation going while the work finishes behind the scenes.
Extended Thinking reasons and speaks at once
The defining trait of Extended Thinking is that reasoning and speech happen in parallel.
Instead of going quiet while it works, the model opens with a verbal cue along the lines of "Let me check that," then narrates its progress as it moves through a multi-step task. Anyone who has waited on a phone call while the other party silently looks something up will recognize the problem it addresses.
Google's demonstrations include building React components from hand-drawn sketches plus spoken feedback, coordinating multiple bookings asynchronously without interrupting the conversation, and producing business plans and marketing toolkits through speech alone.
Where it lands on the benchmarks
The numbers favor Extended Thinking.
It scored 82.6 on Artificial Analysis' Speech to Speech Quality Index, taking the top overall spot. On agentic task completion it reached 68.6 percent on τ-Voice and 35.1 percent on Sierra's τ-Voice-banking benchmark. On Big Bench Audio, which measures reasoning over audio, it recorded 97.7 percent. Google says it reached those levels while staying price-competitive against other frontier models.
Gemini 3.8 Live placed second in the Speech Agent Arena, where rankings come from user preference. On ServiceNow's EVA-Bench, a benchmark for voice agents, Google says both models pushed the frontier by balancing accuracy against conversational quality. The EVA-Bench figures were measured on the Live API running on the Gemini Enterprise Agent Platform.
Benchmark results depend on how the evaluation is designed, so they do not translate directly into how good a model feels in use. Still, claiming top-tier placement without giving up the price point is hard for any company trying to put voice agents into production to ignore.
Developer support and safety measures
Through the Gemini Live API, developer platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents support the new models. Those platforms absorb the awkward parts of real-time media streaming, which leaves developers free to focus on the user experience. Google named Salesforce, Genspark and Lumeris among the companies keen on the release.
On safety, every piece of audio the models generate carries a SynthID watermark. It is woven into the audio itself and is imperceptible to human ears, so AI-generated audio remains detectable after the fact. The better speech synthesis gets, the more weight that kind of detection carries.
Availability
Gemini 3.8 Live is available to developers through the Gemini API and Google AI Studio. Enterprises get it as a private preview in Gemini Enterprise, with Gemini Enterprise for Customer Experience to follow. General users can reach it in Search Live.
Gemini 3.8 Live Extended Thinking is likewise available to developers through the Gemini API and Google AI Studio. Enterprise access begins as a private preview in Gemini Enterprise, followed by Gemini Enterprise for Customer Experience and Google Workspace business customers. For general users it appears in Gemini Live, in Workspace Docs for Google AI Pro and Ultra subscribers, and in Gmail and Keep for all Google AI subscribers.
Summary
Gemini 3.8 Live covers scale and cost efficiency while Gemini 3.8 Live Extended Thinking covers reasoning, forming a two-model voice AI lineup. Automatic switching across 97 languages, near real-time visual understanding and background tool execution all work to keep conversations from stalling, and Extended Thinking took the top overall spot on the Speech to Speech Quality Index with 82.6. Developers can start from the Gemini API and Google AI Studio, while general users reach the models through Search Live and Gemini Live.
