Google Drops Gemini 3.8 Live: Real-Time Voice Agents That Actually Think
Google just released Gemini 3.8 Live and Extended Thinking. Voice agents can now reason, run background APIs, and talk at the same time.
Models · Source: Hacker News
What happened
Google just launched Gemini 3.8 Live and 3.8 Live Extended Thinking. These are near real-time voice models built for fluid dialogue and complex reasoning. They handle visual inputs on the fly to enrich conversations with immediate context. The models automatically detect and switch between 97 supported languages mid-sentence. You can access them today through the Gemini API and Google AI Studio.
The Extended Thinking model is the heavy hitter for enterprise tasks. It talks and reasons at the exact same time. It uses verbal cues like Let me check that while running multi-step background tasks. This keeps the user engaged with live progress narration instead of dead silence. It dominates benchmarks, hitting number one on the Speech to Speech Quality Index and scoring high on agentic task completion.
The standard 3.8 Live model focuses on scale and cost efficiency. It executes tool and API calls in the background without breaking the conversation flow. The model acknowledges requests and keeps chatting while tasks finish. Google is pushing this hard into their ecosystem. They are integrating it directly into Workspace apps like Docs, Gmail, and Keep, alongside Search Live and enterprise platforms.
Key facts
- 97 — supported languages the model can switch between mid-conversation
- 82.6 — score on Artificial Analysis Speech to Speech Quality Index for Extended Thinking
- 68.6% — score on agentic task completion on Sierra Voice benchmark
- 97.7% — score on Big Bench Audio for reasoning capabilities
Why it matters
Voice interfaces are finally moving past the awkward walkie-talkie phase. Until now, building a voice agent meant dealing with dead air while the model waited for an API response. Gemini 3.8 Live solves this by running asynchronous function calls in the background while keeping the chat going. You can now build voice apps that feel like talking to a competent human assistant. The agent can type on a keyboard, query a database, and speak to you all at once. This completely changes how users interact with software.
This raises the baseline for customer service and enterprise automation. With models natively supporting 97 languages and real-time visual context, the barrier to building global, multimodal agents just collapsed. Companies relying on basic text-to-speech wrappers will get wiped out. Native speech-to-speech models that narrate their own progress and handle interruptions gracefully are the new standard. Developer platforms like LiveKit, LangChain, Pipecat, and Vercel are already supporting the Live API. This means the infrastructure is ready for immediate deployment by small teams.
For builders
Build asynchronous voice agents without dead air
Gemini 3.8 Live executes API calls in the background while maintaining conversation. You can build complex workflows where the AI acknowledges the request and chats while waiting for the database. Customers pay for fluid experiences, and clunky voice bots will lose market share rapidly.
Leverage native real-time visual context
The models process visual inputs in near real-time during voice chats. You can build onboarding tools or coding assistants that look at a user screen and talk them through it. Founders building multimodal support tools have a massive new capability here to capture enterprise budgets.
Deploy instantly via existing developer platforms
Google partnered with platforms like LiveKit, LangChain, and Vercel to manage the complex media streaming infrastructure. You can focus entirely on the user experience instead of wrestling with WebRTC protocols. Engineering teams save months of infrastructure work and can ship production-ready voice agents today.
My take
I have been waiting for voice models to handle background tasks without freezing. Voice is the new terminal, and Google just gave us a much better command line. The ability to run background API calls while the model buys time with natural filler words is exactly how humans work. If you are still building text-only chatbots in 2026, you are already obsolete.
Original reporting: Hacker News. This is my rewrite and opinion.