OpenAI releases GPT-Live-1 for voice agents

OpenAI has released GPT-Live-1 in its API, a voice model that listens and speaks at the same time to create more natural assistants for applications, calls, and enterprise workflows.

The main difference is how it handles a conversation. Instead of waiting for you to finish, converting your voice to text, processing it, and generating a spoken response, GPT-Live-1 analyzes incoming and outgoing audio within a single model.

That lets it react more effectively when you interrupt it, pause, or change your mind halfway through a sentence. On a call, for example, you can ask to make a reservation, correct the date while the assistant is responding, and continue without starting over.

Fewer interruptions and more natural conversations

Traditional voice systems usually combine three parts: a tool that converts speech to text, a model that decides what to say, and another that turns the response into speech. Each connection can add delay and cause the agent to lose context.

GPT-Live-1 brings listening and speaking together in a single model. It can also handle silence, background noise, and brief confirmations without narrating every step or cutting off the conversation unnecessarily.

In early evaluations by Speak, a language learning platform, students had more time to think and interruptions fell by nearly 80% compared with earlier turn-based systems. OpenAI presents this as an early result, not a guarantee for every application.

OpenAI's evaluations also place GPT-Live-1 30 percentage points above GPT-Realtime-2.1 on Full Duplex Bench, a test focused on conversations where both sides can speak at the same time. When combined with GPT-6 Astra at a medium reasoning level, it took first place on Tau3, an evaluation of complete tasks for voice agents.

A voice model that delegates complex work

GPT-Live-1 does not have to handle every part of a task on its own. It can delegate deeper reasoning and tool calls to a connected text model running in the background, such as GPT-6 Astra, or to third-party models.

This lets a company choose different combinations depending on the job:

• A fast, affordable model for updating orders or scheduling appointments.
• A model with greater reasoning capacity for solving complex customer problems.
• In-house tools for checking inventory, internal systems, or calendars.

While that work is happening, the agent can keep the conversation going. You can also define its tone, speed, and style through instructions instead of being limited to a fixed voice and default behavior.

From calls to everyday assistants

The API includes automatic speech recognition transcripts, response text, letter and number comprehension, and settings that prioritize keywords. Although it does not work as an exclusively turn-based system, it includes turn detection so developers can maintain clear boundaries when needed.

It also supports full-duplex voice agents on phone calls. Possible uses include:

• Restaurant reservations.
• Customer service and support.
• Order updates.
• Appointment management.
• Escalating cases to a person when the agent cannot resolve them.

OpenAI says it is expanding its voice selection to include more accents, dialects, and languages. The company expects to add more options over the coming months.

What changes for you

If you use an app with a voice assistant, the most noticeable improvement will not necessarily be that it responds faster, but that the conversation feels less rigid. You will be able to talk over it, correct yourself, pause, or continue with noise around you without repeating everything from the beginning.

For developers, the update reduces some of the work needed to coordinate speech recognition, reasoning, and audio synthesis. GPT-Live-1 is available in the API for $0.05 per minute for the front-end voice layer. The reasoning model, tools, and infrastructure connected behind it are chosen separately.

The next thing to watch will be how it performs in long conversations and real calls, where both accuracy and cost matter. Voice is no longer just a way to dictate commands to an AI. It is starting to become an interface that can listen, act, and hand control to a person when needed.