OpenAI Launches GPT-Live-1 API for Full-Duplex Real-Time Voice Agents
OpenAI has officially launched the GPT-Live-1 model in its API, allowing developers to build full-duplex, real-time voice applications with native interruption handling, tool calling, and telephony integration. The unified model processes speech input and output directly in one step, replacing traditional multi-stage pipelines. By bypassing separate Speech-to-Text, LLM processing, and Text-to-Speech steps, GPT-Live-1 dramatically reduces latency and architectural complexity for voice applications. Early adopters reported reducing codebases by up to 80% and decreasing accidental interruptions during user pauses by nearly 80%. Pricing for the front-end voice layer is set at $0.05 per minute (roughly $3.00 per hour), with backend reasoning and tool calls billed separately. GPT-Live-1 scored 30 percentage points higher than GPT-Realtime-2.1 on Full Duplex Bench, ranked #1 on Tau3 benchmark tasks when paired with GPT-6 Astra, and introduced 12 new voice options along with keyword biasing support.
## BACKGROUND
Traditional voice AI agents rely on cascaded pipelines combining speech recognition (ASR), a text-based LLM, and text-to-speech synthesis (TTS), which creates noticeble response lag and unnatural turn-taking. Native speech-to-speech models unify speech perception and audio generation, enabling full-duplex communication where both parties can listen and interrupt simultaneously while preserving vocal nuances.