Skip to main content

OpenAI's New Voice API: Talk, Interrupt, and It Won't Miss a Beat

OpenAI Opens Its Voice Engine to Developers

OpenAI has quietly made the voice technology behind ChatGPT available to developers through a new API called GPT-Live-1. The move means any developer can now build full-duplex voice assistants—apps that listen to users and respond in real time, just like talking to a person. The front-end voice processing costs $0.05 per minute, while any heavy reasoning gets billed separately based on the back-end model you choose.

Image

Listen and Speak at the Same Time

Traditional voice AI works in three clunky steps: speech recognition, language model reasoning, then speech synthesis. The system waits for you to finish before it even starts thinking about a reply. GPT-Live-1 throws that pipeline out the window. It generates voice output while still listening, so it can handle interruptions, pauses, agreement, and fast back-and-forth conversation naturally. It even decides when to keep talking, when to pause, when to listen, or when to call a tool.

OpenAI says this architecture cuts the delay that makes voice interactions feel robotic. Developers no longer have to stitch together multiple independent models. In testing, GPT-Live-1 beat GPT-Realtime-2.1 by 30 percentage points on the Full Duplex Bench, with big gains in turn-taking delay and interactive behavior. When paired with GPT-6 Astra at medium reasoning intensity, it also took first place in the Tau3 test, which measures end-to-end voice agent tasks.

Complex Thinking Goes to the Back-End

GPT-Live-1 doesn't try to do everything. OpenAI splits the work: the voice model handles real-time listening and speaking, while a back-end model handles deep reasoning. When you ask something that needs searching, complex analysis, or a longer task, GPT-Live-1 hands it off to the back-end model and keeps the voice conversation flowing. That means developers can pick the right back-end for the job—a fast, cheap model for simple tasks, or a stronger reasoner for tough questions.

In practice, GPT-Live-1 supports phone scenarios natively, so it's ready for full-duplex voice agents like restaurant reservations or customer support. Early tester Speak, a language learning platform, found that GPT-Live-1 cut the system's active interruptions of learners by nearly 80% compared to older round-based systems. The model also provides native voice recognition transcription and response text, supports keyword bias and turn detection, and is optimized for noisy environments. OpenAI expanded the available voices to cover more accents, dialects, and languages. Developers can even combine GPT-Live-1 with tools like Codex—the voice model takes the task, the back-end tools process it, and the voice conversation picks right back up.

Key Points

  • GPT-Live-1 API is now open to developers at $0.05 per minute for front-end voice processing.
  • Full-duplex architecture lets the model listen and speak simultaneously, handling interruptions and pauses naturally.
  • Complex reasoning is offloaded to a separate back-end model, so developers can mix and match based on cost and need.
  • Early tests show a 30-point lead over GPT-Realtime-2.1 on the Full Duplex Bench and an 80% reduction in interruptions for language learners.
  • Native phone support, expanded voices, and noisy-environment optimization make it ready for real-world voice agents.