OpenAI released the API for its voice model "GPT-Live-1" on 2026-09-10, enabling more precise control over the speaking style and behavior of voice agents.

GPT-Live-1 employs a "full-duplex" system, allowing it to listen and speak simultaneously. OpenAI explained that unlike traditional concatenated architectures—which combine speech recognition, large language models, and speech synthesis—this single-model approach for processing both audio input and output reduces latency in areas such as interruption handling.

A key feature is the ability to delegate resource-intensive tasks, such as inference and tool calls during a conversation, to a backend text model. This allows the agent to maintain the flow of conversation even while performing complex tasks. Developers can freely configure the agent's tone, speed, and conversational style via system prompts.

The API is priced at $0.05 per minute for the frontend voice layer. OpenAI stated that it intends to expand supported voice options and languages over the coming months.

Source: