OpenAI opened GPT-Live-1 in its API on September 10, 2026, extending the full-duplex voice model first shown in ChatGPT to third-party apps and business workflows. Unlike cascaded speech-to-text, language-model, and text-to-speech stacks, GPT-Live-1 listens and speaks in one model so interruptions, pauses, and mid-turn course changes stay in the conversational flow.
The API release emphasizes developer control. Teams can shape tone, pace, and style through the system prompt, expand across a wider set of voices and languages, and keep ASR transcripts plus response text for product UI. For deeper work—search, booking logic, complex support—GPT-Live-1 can hand off reasoning and tool calls to a backend text model such as GPT-6 Astra or a third-party model while the spoken session continues. OpenAI also lists telephony support for phone-based agents, from reservations to customer care.
On OpenAI’s evaluations, GPT-Live-1 improves Full Duplex Bench scores by 30 percentage points over GPT-Realtime-2.1, with sharper turn-taking latency and interactive behavior. Paired with GPT-6 Astra at medium reasoning effort, the company says the stack ranks first on Tau3, its end-to-end voice-agent intelligence benchmark. Early design partners echo the architecture shift: Speak reported nearly 80% fewer interruptions during learner thinking pauses versus prior turn-based tutors, and other customers cite simpler codebases and more natural call handling.
Pricing for the front-end voice layer is $0.05 per minute, billed per second; backend model and tool usage are separate. Documentation covers WebRTC browser sessions, WebSockets, and managed Responses or client-managed delegation. DigiEditorial verified the launch against OpenAI’s product post and independent coverage; availability is live for API developers today.