OpenAI's new voice API listens and speaks at the same time, accepts interruptions and delegates tasks. See architecture, metrics and precautions for voice service.
Direct answer
On September 10, 2026, OpenAI released the GPT-Live-1 API, a full-duplex voice model capable of processing incoming and outgoing audio simultaneously. For companies, the change can reduce latency and simplify service architectures, but production still requires consent, data protection, accent testing, human escalation and resolution measurement — not just conversational naturalness.
What changes with full-duplex
In traditional systems, speech recognition, reasoning and speech synthesis operate in stages. The GPT-Live-1 processes audio input and output in a single model, allowing the person to interrupt, hesitate or change direction without waiting for a hard turn to finish.
Voice and reasoning can be separated
The voice layer can delegate reasoning and tool usage to a text model in the backend. This separation allows you to choose cost and depth per task, keeping the conversation fluid as authorized systems query data or perform actions.
What to measure in customer service
Time until the first response is not enough. Measure incorrect interrupt rate, intent completion, handoffs, repeatability, satisfaction, cost per contact, and accuracy of each action. Compare the pilot to the current channel using the same set of cases.
Privacy and security remain central
Report when people speak to AI, limit retention and access, remove unnecessary data, and record consent where applicable. Material actions must require confirmation, an audit trail, and a rapid path to human assistance.
Nexus Reading
Natural voice brings AI closer to routine, but trust comes from the complete design of the experience. Tone, pause, interrupt, context, allow, and crash recovery need to work as a single measurable service.
FAQ
What does full duplex voice mean?
It means the system can listen and talk simultaneously, handling interruptions and conversational signals in real time.
Does GPT-Live-1 perform actions on its own?
It can delegate use of tools to a backend, but permissions, confirmations and limits are the responsibility of the application.
What is the main metric?
The correct completion of the intention, with acceptable safety and cost, is more important than the sensation of speed alone.
Essential guides to delve deeper into the decision
This editorial analysis was produced by Nexus from the official sources below, consulted on September 11, 2026. The text is original and interprets practical implications for companies.
- OpenAI — Build more natural voice experiences with GPT-Live-1 in the API: official launch, capabilities, reviews, price and availability.
- OpenAI API — Getting started with GPT-Live: architecture, sessions and official technical guidance for integration.
Date reported by the main source: September 10, 2026.