Google Releases Gemini 3.1 Flash Live: Real-Time Voice AI for Agents in 90+ Languages

Google shipped Gemini 3.1 Flash Live into developer preview on March 26, and the model does something the previous generation couldn’t: it processes audio natively, end to end, without converting to text first.

The old architecture — transcribe speech to text, feed it to a language model, synthesize the response back to audio — introduced latency at every handoff. Gemini 3.1 Flash Live eliminates those seams. It handles audio, images, video, and text within a single 128K token context window and outputs up to 64K tokens of audio and text per call. Real-time streaming runs over WebSockets for full-duplex communication, which means the model can interrupt and be interrupted the way humans actually talk.

Numbers that matter

On ComplexFuncBench Audio — a benchmark measuring multi-step function calling based purely on audio input — the model scored 90.8%. That’s not a reading-comprehension score with audio as a delivery mechanism. It means the model can hear “book me a table for four at 8pm, but not at the same restaurant as last week” and correctly chain the tool calls to execute it.

Context window for conversation history doubles what was available in previous Flash Live models. For voice agents that need to hold context across a long session — customer service, interactive tutoring, real-time code review — that’s a practical constraint that just got removed.

The model supports over 90 languages natively. Google is positioning it as the foundation for voice-based AI agents, not just a voice interface bolted onto a text model.

Search Live goes global

Alongside the API release, Google expanded Search Live to every country and territory where AI Mode is available — more than 200 countries and territories now have access to camera-based visual search powered by voice AI. Point your phone at something, ask a question out loud, and get an answer that understands both what you’re looking at and what you’re saying.

What this means for developers

Gemini 3.1 Flash Live is available through the Gemini API and Google AI Studio starting today. Developers building voice agents have historically faced a choice between quality and latency — low-latency systems sounded robotic, high-quality systems introduced noticeable lag. A native audio model with strong benchmark scores and 200+ country reach changes that calculus.

The agent tooling ecosystem — function calling, tool use, agentic workflows — all carry over from existing Gemini 3.1 Flash integrations. The jump to Flash Live is an API model swap, not a rewrite.

For anyone building conversational AI products, this is the most important release from Google so far this year.