OpenAI Releases Bidirectional Voice Model 'GPT-Live 1' on AI Gateway

OpenAI has released a new voice model, GPT-Live 1, on the AI Gateway platform. This model applies full-duplex technology, moving away from the turn detection method used by previous voice models, to provide a natural conversational experience where users can interrupt or add details while the AI is speaking.

One of the core features of GPT-Live 1 is client delegation. This allows the voice model and background text models to be configured independently. By default, a session runs only the voice model, but when complex tasks are required, they can be delegated to other text models on the AI Gateway for processing. For example, GPT-Live 1 can maintain the conversation while a text model like openai/gpt-5.6-sol performs tasks in the background, returning the results to the voice channel to be spoken.

Developers must install AI SDK 7 and @ai-sdk/openai 4.0.67 or later and integrate via a WebSocket client. Requests to delegated models are billed separately through the AI Gateway, and voice session usage continues to accumulate even while delegated tasks are being executed.