# GPT-Live-1 allows for real-time voice interruptions at $0.05 per minute > Source: > Published: 2026-09-10 23:54:08+00:00 # GPT-Live-1 allows for real-time voice interruptions at $0.05 per minute The GPT-Live-1 API is finally here for developers, and the biggest shift is the move to full-duplex communication. Unlike the old "push-to-talk" or sequential turn-taking patterns, this model can listen and speak simultaneously. In interactivity benchmarks, it hit 80.1%, which is a massive jump from the 45.4% seen in previous iterations. This means the latency and "interruption" handling actually feel natural instead of like a choppy walkie-talkie conversation. ## How does the duplex flow actually work? Most voice AI setups follow a linear path: Speech-to-Text (STT) -> LLM processing -> Text-to-Speech (TTS). That pipeline is exactly why there's always that awkward two-second pause before an AI responds. GPT-Live-1 collapses this. Because it's full-duplex, the model isn't just waiting for a "silence" trigger to start processing; it's constantly monitoring the audio stream. If you try to interrupt the AI while it's speaking, it can actually detect your voice over its own output and stop immediately to listen. This is the specific area where that 80.1% interactivity score comes from. For anyone building a voice assistant or a real-time tutor, this removes the need for complex "interrupt" logic on the client side. ## The cost of real-time latency The price point is the main hurdle. At $0.05 per minute, the costs scale quickly if you're building a consumer-facing app. - **10-minute conversation:** $0.50 - **1 hour of usage:** $3.00 ## When should you avoid using it? Don't use GPT-Live-1 if your app doesn't actually need "live" feel. If you're building a voice-memo summarizer or a simple Q&A bot where a 2-second delay doesn't matter, sticking to the standard Whisper (STT) and TTS pipeline is significantly cheaper and easier to manage. Full-duplex is specifically for scenarios where the *cadence* of the conversation is the product, such as language learning or live customer support. ## Getting started with the integration To implement this, you'll be dealing with WebSocket connections rather than standard REST API calls because the audio needs to stream continuously. You'll need to handle audio encoding—typically PCM 16-bit at 24kHz—to ensure the model captures the nuance of the speech without introducing artifacts. If you're testing this out, I recommend starting with a small set of specific prompts to see how it handles "barge-ins" (when the user speaks over the AI). If the model doesn't stop instantly, check your audio buffer sizes on the client side, as that's usually where the lag originates, not the API itself. [Next Nemotron 3 Ultra hits 2.5x higher concurrency with full-stack NIM optimizations →](/en/threads/9167/) ## All Replies （3） Finally! I'm sick of the awkward silence during my current API calls. Does this actually fix the 200ms lag on Twilio? I want to try this tonight for my app. I'm curious if it handles 40ms latency better than Vapi. I'm dying to test this tonight. Does it play nice with WebSockets or is there a specific SDK for the 0.05 rate?