OpenAI has released GPT‑Live‑1 directly in the API so developers can build voice-enabled apps and business workflows with the same natural spoken model that powers ChatGPT. The model can listen and speak at the same time (full-duplex), and it can delegate deeper reasoning and tool actions to backend models such as GPT‑6 Astra.
Why it matters for voice agent builders
Traditional voice agents chain speech-to-text, a reasoning model, and text-to-speech. Every handoff adds latency and breaks the natural rhythm of a conversation. GPT‑Live‑1 handles listening and speaking in a single model, responding to interruptions and acknowledgements in real time while work happens in the background. Developers can also shape an agent’s tone, pace, and conversational style through the system prompt.
Key strengths announced
- Interruption handling: a single model reasons over incoming and outgoing audio together, avoiding brittle STT–LLM–TTS handoffs. In early evaluations, the language-learning app Speak saw interruptions during thinking pauses cut by nearly 80% versus previous turn-based systems.
- Reasoning and tool-call delegation: GPT‑Live‑1 can push deeper reasoning and tool calls to a backend text model such as GPT‑6 Astra or a third-party model.
- Tone, pace, and style control through the system prompt.
- Silence, context, and background noise: better handling of noisy environments without over-narrating every step.
- Long-session reliability for extended conversations.
- Telephony support: full-duplex voice agents for phone calls, from restaurant reservations to customer support.
Real-world results
- Yelp reports improved turn-taking and call handling in Yelp Host and Hatch, with callers speaking fuller, more natural sentences.
- A health-tech CTO cited an 80% smaller codebase (23K lines removed) versus a cascaded build, enabling natural real-time patient conversations.
- On benchmarks, GPT‑Live‑1 improved Full Duplex Bench performance by 30 percentage points over GPT‑Realtime‑2.1 and, paired with GPT‑6 Astra at medium reasoning effort, ranked #1 on Tau3 for frontier voice-agent intelligence.
- Cognition (Devin) shows voice collaboration with an AI engineer feeling like working alongside a teammate.
New voices and pricing
OpenAI expanded the voice catalog across accents, dialects, and languages (Quartz, Ripple, Vesper, Willow, Stone, Gleam, Meridian, Bossa, Tempo, Beacon, Delta, Cinder, and more), with additional options planned. GPT‑Live‑1 is available today in the API at $0.05 per minute for the front-end voice layer. Enterprises can also build on it through OpenAI Presence for trusted, action-taking agents.
Suijin Greenverse view
Voice agents are moving from stop-start chatbots toward natural phone-grade conversations. The interesting architecture shift here is not the voice itself — it is the delegation boundary: the conversation engine stays fast and human, while real reasoning and actions happen behind it. Teams building customer-facing voice workflows should evaluate interruption handling first; it is where the difference is most visible.
Source: OpenAI, September 10, 2026