OpenAI's full-duplex voice model has been open to developers since September 10: speaking latency divided by almost two, billing per minute of audio.
OpenAI opened GPT-Live-1 to its API on September 10, 2026, charged $0.05 per minute for the voice layer alone. The model processes incoming and outgoing audio in a single system rather than chaining speech recognition, language model, and synthesis, which removes the latency accumulated by this three-step pipeline.
The figures put forward by OpenAI compare the model to its predecessor GPT-Realtime-2.1: 86.2% of tasks passed on the first try on Tau3 Voice Intelligence compared to 45.7%, a speaking latency reduced from 1.41 to 0.798 seconds, and an interactivity score of 80.1% on Full Duplex Bench compared to 45.4%.
A special feature of the architecture is that reasoning and tool calls can be delegated to a model in the background while GPT-Live-1 keeps the conversation going, instead of interrupting the exchange for processing. Yelp uses it for taking reservations and orders, Speak says it has reduced interruptions during its language lessons by nearly 80%, and Fin has integrated it into its customer support.
Read the full article on the original source
Go to the source →