GPT-Live-1 turns full-duplex voice behavior into an API primitive that developers can place inside their own applications. This matters now because a natural ChatGPT voice behavior has become infrastructure that external product teams can test, route, and operate.
OpenAI says GPT-Live-1 is available through the API and can continue listening while a voice agent is speaking. The release also gives developers more control over how a voice agent speaks and acts during a conversation. The mechanism is one continuous session that coordinates incoming audio, generated speech, interruptions, conversational state, and tool-driven actions without forcing a rigid turn boundary.
Voice-agent evaluations should cover interruption recovery, latency, tool transitions, consent, and the state retained after a user changes direction. The evidence is OpenAI's launch material; it does not yet establish independent reliability, cost, or safety across noisy production environments. Watch whether deployed applications publish measurable interruption accuracy, end-to-end latency, and failure recovery under real conversational load.