OpenAI has opened access to its GPT-Live-1 voice model via API, allowing developers to integrate real-time, full-duplex voice capabilities into their applications. The model is designed to handle interruptions, pauses, and natural back-and-forth dialogue without requiring users to wait for the system to finish speaking.
The pricing for the GPT-Live-1 voice frontend is set at $0.05 per minute. This cost covers the real-time audio processing, while any additional complex reasoning tasks performed by backend models are billed separately according to their specific pricing structures.
The architecture separates real-time interaction from deep reasoning. Developers can pair GPT-Live-1 with various backend models, allowing them to balance performance, reasoning capabilities, and costs based on the complexity of the task, such as simple appointment scheduling versus complex analysis.
In performance tests, GPT-Live-1 outperformed the GPT-Realtime-2.1 model by 30 percentage points in the Full Duplex Bench evaluation. It also secured first place in the Tau3 benchmark, which assesses an agent's ability to complete end-to-end tasks.
Early adopters include platforms like Speak, which reported an 80% reduction in unwanted system interruptions during language learning sessions. The model is optimized for noisy environments and supports telephone-based applications, making it suitable for customer service and automated booking systems.
