OpenAI introduced GPT-Live-1 in its developer API on September 10, according to the OpenAI announcement. The model can listen and speak simultaneously, supporting native turn detection, automated speech recognition transcripts, and text output. System prompts allow developers to tune the voice layer's tone, pace, and style across an expanded set of voices, accents, and languages.
The architecture splits conversational audio from background execution. According to the OpenAI Live developer guide, GPT-Live-1 delegates tasks either directly to a supported model through Responses delegation or to external services, models, and agents via client delegation. Integration options include WebRTC for browser sessions and WebSockets for server environments, where trusted servers must hold the API key.
The separation also affects billing and control:
- Billing boundaries: The front-end voice stream costs $0.05 per minute and is billed per second, as detailed in the OpenAI announcement and developer guide. Backend model and tool usage is billed separately; the voice rate is not an all-inclusive agent price.
- State and cancellation: As documented in the developer guide, the application manages permissions, confirmations, private function execution, and durable task state. Interrupting speech does not automatically cancel backend work.