OpenAI has opened API access to its GPT-Live-1 model, which supports full-duplex voice interaction. This allows the AI to listen and speak simultaneously, enabling more natural communication compared to previous systems that required strict turn-taking.
Developers can now integrate GPT-Live-1 into their applications to customize voice interaction scenarios. The model is designed to handle interruptions, which has already been tested by the language learning service Speak, resulting in an 80% reduction in interruptions during sessions.
In the Tau3 benchmark for customer service tasks, GPT-Live-1 achieved a score of 86.2%, significantly outperforming GPT-Realtime-2.1 and GPT-Realtime-2, which scored 45.7% and 42.4% respectively. The model also shows a 30% improvement in the Full Duplex Bench.
The model is already in use by Yelp Host to facilitate restaurant reservations. While GPT-Live-1 handles the dialogue, backend models like GPT-6 Astra or other variants manage data retrieval and task execution.
Access to GPT-Live-1 is priced at $0.05 per minute, excluding the costs for the backend model. It is also available through the OpenAI Presence platform for enterprise agent development.
