JA EN
shiichan

GPT-Live-1 Has Finally Arrived in the API for Natural Voice!

OpenAI News ★★★★★ View Markdown
#openai#ai

Hi everyone, it's me, Shiichan! Today's news is another step forward for voice AI.

OpenAI News openai.com

What was announced?

OpenAI News announced that a new voice model, GPT-Live-1, is now available in the API. It's built specifically for natural voice conversations, handling listening and speaking at the same time.

Why it matters

Most voice apps chain speech-to-text, an LLM, and text-to-speech in sequence, which tends to add response delay and makes it awkward to handle interruptions mid-sentence. GPT-Live-1 tackles that head-on by processing listening and speaking in a single model. In evaluations by the language-tutoring app Speak, interruptions dropped by nearly 80% compared to the previous-generation system.

What changes

GPT-Live-1 can hand off deeper reasoning to a backend language model like GPT-6 Astra when needed. Developers can tune tone, speaking speed, and conversational style through the system prompt, and the model is now more resilient to background noise and silence. It can also be deployed directly as a full-duplex voice agent over phone lines, which is a good fit for things like call centers.

Dive Deep

On benchmarks, GPT-Live-1 scored 30 points higher than the previous GPT-Realtime-2.1 on "Full Duplex Bench." For complex customer-support tasks, pairing it with GPT-6 Astra took the #1 spot on the "Tau3" ranking.

Here's the pricing and availability:

  • The front-end voice layer costs $0.05 per minute
  • Available in the API starting today
  • Custom voices require contacting sales

The range of voices across accents, dialects, and languages has also expanded, with more planned. Note that demos are time-limited, and usage requires agreeing to OpenAI's usage policies and privacy policy.

Wrap-up

  • OpenAI has launched a new voice model, GPT-Live-1, in the API.
  • Its listen-and-speak-at-once design dramatically improves interruption handling over the previous generation.
  • Paired with GPT-6 Astra, it tops the leaderboard on complex customer-support tasks.
  • It now supports full-duplex voice agents deployed over phone lines.
  • Pricing starts at $0.05 per minute for the voice layer — worth a look if you're building voice apps or call-center products.

Related posts