Google DeepMind ships Gemini 3.1 Flash Live, the firm’s highest-quality audio model built for real-time dialogue with tighter latency and sharper tonal awareness. Developers tap the Gemini Live API in Google AI Studio, enterprises fold it into customer workflows, and consumers access upgrades through Search Live across over 200 countries. Better rhythm and reasoning let voice agents complete multi-step tasks without breaking conversational flow.

Performance jumps stem from deeper training on complex instruction following and long-horizon reasoning amid interruptions typical of live speech. On ComplexFuncBench Audio, the model scores 90.8%, topping its predecessor, while Scale AI’s Audio MultiChallenge records 36.1% with thinking enabled. Watermarking via SynthID tags every output to curb misinformation, and Verizon, LiveKit and The Home Depot report smoother, more natural exchanges.

Gemini 3.1 Flash Live sustains threads for twice as long as earlier versions and adapts dynamically to frustration or confusion without losing intent. Multilingual capability drives global reach while preserving low latency in noisy settings. All audio from 3.1 Flash Live carries imperceptible provenance marks for reliable detection of synthetic origin.