OpenAI Launches GPT-Live-1 API With Full-Duplex Voice, Custom Voices and Telephony Support

OpenAI Brings GPT-Live-1 to Developers With Natural Voice Conversations at $0.05 Per Minute

The new model is designed around real-time interaction, allowing users to interrupt an AI assistant, change their question, pause while thinking and continue speaking naturally. OpenAI is also giving developers more control over voice style, background noise handling, tool delegation and telephony applications.

The API release could be particularly important for companies building customer-service agents, tutors, voice assistants, phone-based AI systems and other applications where natural conversation matters.

OpenAI says GPT-Live-1 is available in the API now at $0.05 per minute for the front-end voice layer, while backend models and tools are billed separately.

What Is GPT-Live-1?

GPT-Live-1 is OpenAI's voice model for building real-time conversational applications.

The model was first introduced as part of ChatGPT's newer voice experience. With the API release, developers can now use the underlying voice capabilities in their own applications and workflows.

The biggest technical difference is its full-duplex architecture.

In a traditional voice assistant, several separate components may handle speech recognition, language-model reasoning and text-to-speech generation. These separate stages can introduce delays and make interruptions difficult to handle smoothly.

GPT-Live-1 is designed to handle incoming and outgoing audio together.

That allows the model to better understand the flow of a conversation instead of treating every spoken sentence as an isolated request.

GPT-Live-1 Can Listen While It Speaks

One of the most important features of GPT-Live-1 is simultaneous listening and speaking.

In a normal chatbot interaction, users generally wait for the system to finish before providing the next instruction. Real conversations do not work that way.

People interrupt.

They change their minds.

They say short acknowledgments.

They pause to think.

They add information halfway through a sentence.

GPT-Live-1 is designed to handle these behaviors more naturally.

OpenAI specifically highlights interruption handling as one of the model's key strengths. The company says the model can reason over incoming and outgoing audio together, reducing the need for multiple handoffs between separate speech systems.

This could make a significant difference for applications where conversation speed and natural turn-taking are important.

Developers Can Delegate Deeper Reasoning to Other Models

GPT-Live-1 does not necessarily have to perform every complex task itself.

OpenAI designed the system so developers can pair the voice layer with backend models and tools.

For example, GPT-Live-1 can handle the spoken interaction while a backend model such as GPT-6 Astra performs deeper reasoning. Developers can also connect their own tools and agent frameworks.

This creates a separation between conversation and execution.

The voice model can concentrate on keeping the interaction natural, while a backend system handles tasks that require deeper reasoning, information retrieval, coding or other tools.

That architecture could be useful for enterprise AI systems where a voice interface needs to interact with existing business software.

New Controls for Voice Personality

OpenAI is also giving developers more control over how GPT-Live-1 sounds and behaves.

The company says developers can shape the assistant's:

Tone
Pace
Conversational style
Response behavior
Language
Overall voice experience

The API release also expands the available voice selection across different accents, dialects and languages.

This matters because voice assistants are increasingly being used as customer-facing products.

A banking assistant, language tutor, healthcare support system or restaurant phone agent may require a very different conversational personality.

Giving developers more control makes it easier to design the voice experience around a particular audience.

Better Handling of Background Noise

Real-world voice conversations rarely happen in perfect environments.

People talk to assistants while walking outside, sitting in cafés, working in offices or dealing with background conversations.

GPT-Live-1 includes improvements for background noise and silence.

OpenAI says the model can better recognize silence and background audio without unnecessarily interrupting the conversation or narrating every internal step aloud.

This is important because a voice agent that constantly mistakes background noise for a user command can quickly become frustrating.

The goal with GPT-Live-1 is to make the interaction behave more like a normal conversation.

GPT-Live-1 Supports Phone-Based AI Agents

Another major addition is telephony support.

Developers can use GPT-Live-1 to build full-duplex voice agents for phone conversations, including applications such as customer support and reservations.

This opens a large potential market.

Businesses already operate call centers for bookings, customer questions, technical support and sales. AI voice agents can potentially automate portions of these conversations while still allowing human employees to handle more complicated cases.

OpenAI's launch announcement specifically points to restaurant reservations and customer support as examples of where GPT-Live-1 can be deployed.

The technology could therefore become particularly important for companies looking to move beyond text-based AI chatbots.

GPT-Live-1 Can Work With AI Tools

Voice interaction becomes more useful when the assistant can actually do something with what the user says.

GPT-Live-1 supports delegation to backend reasoning models and tools, allowing developers to build systems where a spoken request can trigger a larger workflow.

Imagine a customer calling a business and saying they want to change an appointment.

A voice agent could understand the request, pass the required information to a backend system and return the result through the same conversation.

The important distinction is that GPT-Live-1 is not simply converting speech into text and reading an answer aloud.

It is designed as a conversational layer that can work alongside the models, tools and agent infrastructure responsible for completing tasks.

OpenAI Reports Stronger Voice-Agent Performance

OpenAI says GPT-Live-1 performed substantially better than GPT-Realtime-2.1 on its Full Duplex Bench evaluation, with a 30-percentage-point improvement in the company's reported testing.

OpenAI also says GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranked first on its Tau3 evaluation for end-to-end voice-agent intelligence.

These results come from OpenAI's own evaluations, so they should be viewed as vendor-reported performance rather than independent confirmation.

Still, the results demonstrate what OpenAI believes is the main advantage of the architecture: better interactive behavior while maintaining access to deeper reasoning through a backend model.

GPT-Live-1 Pricing

OpenAI has made GPT-Live-1 available through its API at $0.05 per minute for the front-end voice layer.

The company says the voice layer is billed based on usage, while the backend model and tool usage are charged separately.

This distinction is important for developers calculating the total cost of a voice application.

A production voice agent may involve several components:

GPT-Live-1 voice processing
A backend reasoning model
External tools
Database or retrieval systems
Telephony infrastructure
Other application services

Therefore, the $0.05-per-minute figure should not be interpreted as the total cost of operating a complete AI voice agent.

OpenAI Is Targeting Businesses as Well as Developers

The release has implications beyond experimental AI applications.

OpenAI highlights customers and use cases involving companies such as Yelp, Speak, Fin and Cognition. Their examples span restaurant calls, language learning, customer support and AI engineering workflows.

One early evaluation from Speak, cited by OpenAI, found that GPT-Live-1 reduced interruptions by almost 80% compared with previous turn-based systems in the company's language-learning testing.

Again, this is an OpenAI-reported customer evaluation rather than an independent benchmark.

Nevertheless, it demonstrates why interruption handling could become an important differentiator for voice AI.

GPT-Live-1 Could Change How Voice Agents Are Built

The broader significance of the release is architectural.

Developers have traditionally needed to assemble several technologies to build a voice assistant.

A typical system might include:

Speech recognition
A language model
Turn detection
Text-to-speech
Tool calling
Agent orchestration
Telephony infrastructure

Each additional component can introduce latency and create more opportunities for errors.

GPT-Live-1 aims to simplify part of that stack by handling the conversational voice experience directly and delegating more complicated reasoning or actions to the backend system.

That could make sophisticated voice agents easier to build.

What Developers Can Build With GPT-Live-1

The technology could be used across a wide range of applications.

Customer Support

Companies could build AI phone agents capable of answering questions, retrieving information and escalating complicated cases.

AI Tutors

Language-learning applications can use real-time conversations where students can speak naturally instead of waiting for a complete response before continuing.

Restaurant and Booking Assistants

Businesses can create phone agents that handle reservations and other common requests.

Enterprise Voice Assistants

Companies could connect voice agents with internal tools and business systems.

AI Coding Assistants

Developers could discuss a software problem verbally while a backend coding agent handles deeper reasoning and implementation tasks.

These are examples of possible applications, not a guarantee that GPT-Live-1 independently performs every workflow without additional engineering.

GPT-Live-1 vs Traditional Voice AI

The biggest difference can be summarized simply.

Traditional voice AI often works as a sequence.

Listen → transcribe → reason → generate → speak.

GPT-Live-1 is designed around a more continuous conversation.

Listen ↔ speak ↔ interrupt ↔ reason through backend systems ↔ continue.

That does not eliminate the need for other components, but it changes where the conversational intelligence sits.

For developers building highly interactive voice applications, that architectural difference could be more important than a simple improvement in speech quality.

OpenAI's launch of GPT-Live-1 in the API marks another major step toward more natural AI voice applications.

The model brings full-duplex conversations, improved interruption handling, background-noise awareness, customizable voice behavior and telephony support to developers.

Its architecture is also notable because GPT-Live-1 can focus on real-time conversation while delegating deeper reasoning and tool use to a backend model or agent framework.

At $0.05 per minute for the voice layer, the service gives developers a relatively straightforward starting point for experimenting with conversational voice applications, although complete application costs will depend on backend models, tools and infrastructure.

The larger trend is clear: AI assistants are moving away from the old question-and-answer model toward systems that can listen, respond, interrupt, reason and take action during the same conversation.

GPT-Live-1 is OpenAI's latest attempt to make that experience available beyond ChatGPT and directly to developers building the next generation of voice-based AI products.

Frequently Asked Questions

What is GPT-Live-1?

GPT-Live-1 is OpenAI's full-duplex voice model available through the API. It is designed for natural real-time conversations where the AI can listen and speak simultaneously.