Gemini 3.1 Flash Live

Google’s streaming speech-to-speech voice model that replaces the classic STT to LLM to TTS pipeline with a single native speech model. Marketed as Google’s biggest voice upgrade, with lower latency, noise-robust listening, interruption handling and multimodal vision input.

Key facts

• Native speech-to-speech: no intermediate text transcription step, so prosody, sarcasm and stress are preserved into the reasoning layer
• Beats Gemini 2.5 Flash by ~19% on multi-step function calling and outperforms competitor models on the Audio Multi-Challenge benchmark
• Supports multimodal vision — the agent can watch a webcam or share-screen feed and reason about what it sees
• Over 70 supported languages, enabling real-time translation use cases
• Free tier available in Google AI Studio with no API key; paid tier removes the ‘training on your data’ clause and raises rate limits
• Pricing: roughly 14 cents per 10-minute call on the paid tier
• Current limitation: stops speaking during function calls — cannot narrate over tool execution the way a prompted Vapi agent can
• Deployment beyond Google AI Studio requires managing persistent websocket connections — less plug-and-play than ElevenLabs or Vapi for web embedding

Source references

• ^[src-007] Nate Herk cluster — Nate Herk — Voice AI agents cluster (4 videos)

– Videos referenced: Qt3zMBH-FNg