The latest

OpenAI made GPT-Live-1 available through its API on September 10, 2026, bringing the model beyond ChatGPT voice conversations to developers building phone, customer-service, booking and other real-time voice applications. Available immediately at launch, rather than as a closed preview or future release, the model lets users interrupt an agent or redirect a conversation before its reply ends. Its full-duplex design processes incoming and outgoing audio simultaneously, avoiding serial handoffs between speech-to-text conversion, a reasoning model and text-to-speech output.

Details

• Conversation layer: GPT-Live-1 manages conversational pacing, pauses, brief cues and background noise as the voice-facing component of a broader system. A developer can connect it to a text model for deeper reasoning and to tools or an agent framework for actions such as retrieving customer information or completing an authorised request.
• Application control: The application remains responsible for selecting the back-end model, tools and permissions. This separation allows the live voice exchange to continue while another model handles more complex tasks, but interrupting the agent’s speech does not necessarily cancel work already started in the background.
• Integration options: The API automatically transcribes what the model hears and says, while supporting keyword biasing and end-of-turn detection. Connections are available through WebRTC for browser applications, WebSockets for servers, and telephony technologies including SIP.
• Voices and languages: OpenAI added 12 real-time voices whose tone, speed and style can be adjusted through system instructions. The company said the choices span a wider range of accents and languages. Custom voices require contact with its sales team and are subject to separate eligibility conditions.
• Benchmark results: OpenAI said GPT-Live-1 improved its Full Duplex Bench result by about 30 percentage points over GPT-Realtime-2.1. Language-learning platform Speak said early tests produced an approximately 80% reduction in interruptions when learners paused to think.
• Cost structure: The front-end voice layer is priced at $0.05 per minute. Charges for the selected back-end reasoning model, agent tools and required connectivity are calculated separately, making the voice rate the starting point for an application’s total operating cost.

Between the lines

The architecture separates the conversational interface from the model executing the task. An organisation can assign a faster model to simple requests and a deeper model to complex cases without stopping the live exchange. That flexibility makes permission design, action limits, cancellation paths and cost controls central to deployment.

What’s next

Arabic-language teams’ next step is a pre-deployment test covering Modern Standard Arabic and dialect recognition, names and numbers, response quality in noise, and the safety of executed actions. The deployment decision should also review tool logs, permissions and execution limits, then calculate total cost by adding back-end model, tool and connectivity charges to the $0.05-per-minute voice rate.