Skip to main content

OpenAI's new voice API lets developers build real-time talkers for 5 cents a minute

OpenAI's new voice API lets developers build real-time talkers for 5 cents a minute

OpenAI has quietly made a big move: the same voice model that powers ChatGPT's spoken conversations is now available to developers through an API. The model, called GPT-Live-1, lets you build voice assistants that can listen and respond at the same time—no more waiting for the user to finish speaking before the AI chimes in.

And the price? Five cents per minute for the front-end voice processing. That's the part that handles the real-time listening and speaking. If your app needs heavy-duty reasoning, that gets billed separately based on which back-end model you choose.

Image

Listen and speak at the same time

Most voice AI today works like a walkie-talkie: you talk, it waits, then it talks. That means awkward pauses, missed interruptions, and a stilted rhythm. GPT-Live-1 flips that. It uses a full-duplex architecture, meaning it can generate speech while still listening to you. So it handles interruptions, pauses, quick back-and-forth, and even knows when to keep talking, stop, or call a tool.

OpenAI says this cuts down latency dramatically and eliminates the clunky handoffs between separate speech recognition, language, and synthesis models. In tests, GPT-Live-1 beat the previous GPT-Realtime-2.1 by 30 percentage points on the Full Duplex Bench, especially in turn-taking and interactive behavior. When paired with GPT-6 Astra at medium reasoning, it also topped the Tau3 benchmark for end-to-end voice agent tasks.

Complex thinking gets handled by a separate brain

Not every question needs a deep dive. GPT-Live-1 splits the work: it handles the real-time voice interaction, but when you ask something that requires search, analysis, or a longer task, it hands off to a back-end model. That way, developers can pick the right brain for the job—a fast, cheap model for simple stuff, a stronger reasoner for the hard stuff.

This also means you only pay for what you use. The 5-cents-per-minute covers the voice front end; the back-end reasoning is billed at whatever model you choose.

Real-world uses, from reservations to language learning

GPT-Live-1 supports phone scenarios natively, so it's ready for full-duplex voice agents like restaurant booking or customer support. Early tests are promising: language learning platform Speak found that GPT-Live-1 cut the system's interruptions of learners by nearly 80% compared to its previous turn-based system.

The model also provides native speech recognition transcription and response text, supports keyword biasing and turn detection, and is optimized for noisy environments. OpenAI expanded the available voices to cover more accents, dialects, and languages. And developers can combine GPT-Live-1 with tools like Codex—the voice model takes the request, the back-end tools process it, and then the voice conversation picks back up.

It's a pretty clear signal: OpenAI wants developers to build voice-first apps that feel less like talking to a machine and more like talking to a person. And at five cents a minute, it's suddenly a lot cheaper to try.

Key Points:

  • GPT-Live-1 API is now open to developers, priced at $0.05 per minute for front-end voice processing.
  • Full-duplex architecture allows simultaneous listening and speaking, handling interruptions naturally.
  • Complex reasoning is offloaded to separate back-end models, billed per use.
  • Early adopters like Speak saw an 80% reduction in system interruptions.
  • Supports phone scenarios, multiple accents, and integration with tools like Codex.