OpenAI Brings GPT-Live-1 to API, Making Voice Agents More Natural and Responsive

OpenAI Brings GPT-Live-1 to the API for More Natural Voice Agents

OpenAI has launched GPT-Live-1 in the API, giving developers access to the full-duplex voice technology previously introduced through ChatGPT Voice.

The release is designed to make voice agents more natural to interact with while giving developers greater control over how those agents communicate and perform tasks.

OpenAI says GPT-Live-1 can listen and speak at the same time. Rather than forcing users and AI agents into rigid, turn-by-turn exchanges, the model can respond to interruptions, acknowledgements, pauses and changes in conversation as they happen.

The company is positioning the release for voice-enabled applications and business workflows, including phone-based interactions and other applications where conversational timing matters.

Moving Beyond the Traditional Voice-Agent Stack

Traditional voice agents commonly connect several separate components.

A speech-to-text system converts a person's speech into text. A language model processes that text, and a text-to-speech system then converts the response back into audio.

This architecture can work, but each handoff introduces another stage for developers to coordinate.

OpenAI says GPT-Live-1 instead handles incoming and outgoing audio within a single voice model. This allows the model to reason about the conversation while simultaneously listening and speaking.

The model can still delegate deeper reasoning and tool calls to backend models.

This creates a separation between the real-time conversation layer and the deeper reasoning or task-execution layer.

How GPT-Live-1 Handles Real-Time Conversations

The full-duplex architecture is central to the release.

In a conventional turn-based system, the assistant generally waits for the user to finish before generating its response. GPT-Live-1 is designed to continuously process the interaction instead.

That means users can interrupt an agent, change their question, pause while thinking or acknowledge something while the conversation is continuing.

OpenAI specifically highlights interruption handling as a core strength. The company says the approach avoids some of the latency and brittle handoffs associated with chained speech-to-text, language-model and text-to-speech systems.

The model also provides native ASR transcripts and response text, while supporting turn detection for developers who still want explicit conversational boundaries.

GPT-Live-1 Can Delegate Deeper Work

Another important part of the architecture is delegation.

GPT-Live-1 is responsible for maintaining the live conversation, but developers can connect it to backend models for more demanding reasoning or tool-based tasks.

OpenAI says developers can pair the voice layer with models such as GPT-6 Astra or third-party models. This allows teams to choose different reasoning depth, speed and cost characteristics for different tasks.

For example, a voice application could use a faster backend for routine scheduling or order updates while sending more complex customer questions to a model configured for deeper reasoning.

The announcement also demonstrates connecting GPT-Live-1 with Codex, where conversation context can be passed to Codex and its response returned to the live voice interaction.

Developer Control Over Voice Behaviour

OpenAI is also giving developers more control over how their voice agents communicate.

The API supports shaping an agent's:

• Tone

• Pace

• Conversational style

• Interruption behaviour

• Context handling

• Background-noise behaviour

The company says GPT-Live-1 is designed to handle silence and background noise without unnecessarily interrupting the conversation or narrating every internal step.

OpenAI also highlights improved context retention and conversational quality during longer sessions.

This is particularly relevant for applications where conversations cannot be reduced to short question-and-answer exchanges.

Telephony Expands the Potential Use Cases

GPT-Live-1 also supports telephony.

That means developers can build full-duplex voice agents for phone-based interactions, with OpenAI specifically mentioning scenarios such as restaurant reservations and customer support.

This connects the model to a broader category of voice-agent applications where conversations happen over traditional telephone channels rather than only inside websites or mobile applications.

For businesses, the significance is less about voice generation alone and more about integrating conversational AI directly into existing communication workflows.

OpenAI Reports Improvements in Voice-Agent Evaluations

OpenAI reports that GPT-Live-1 improved its performance on the Full Duplex Bench by 30 percentage points compared with GPT-Realtime-2.1.

The company also says that GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranked first on Tau3, an evaluation of end-to-end voice-agent tasks.

These are vendor-reported evaluation results, so they should be interpreted in the context of OpenAI's own testing methodology rather than as an independent industry-wide comparison.

OpenAI also reports an early evaluation from language-learning company Speak, where GPT-Live-1 reduced interruptions during thinking pauses by almost 80% compared with previous turn-based systems.

The company additionally cites experiences from Yelp, Fin and Cognition involving voice interactions, support workflows and AI-assisted engineering conversations.

Voice Options Expand

OpenAI is expanding its selection of real-time voices with GPT-Live-1.

The company says developers have access to a broader selection covering different accents, dialects and languages. It also says that voice options and language availability will continue to expand.

Custom voice access is available through OpenAI's sales process for eligible customers.

Pricing and Availability

GPT-Live-1 is available in the OpenAI API as of September 10, 2026.

OpenAI lists the front-end voice layer at $0.05 per minute. Developers can then pair GPT-Live-1 with the backend model and agent infrastructure appropriate for their application.

This architecture gives developers the ability to separate the cost and capabilities of real-time conversation from deeper reasoning and task execution.

What GPT-Live-1 Means for Voice AI Development

The broader significance of GPT-Live-1 is the shift from voice as an interface layer to voice as an active agent interface.

Instead of treating speech recognition, reasoning and speech generation as completely separate stages, the architecture brings continuous interaction into the core voice model while allowing deeper work to happen through backend models and tools.

For developers, this could simplify parts of voice-agent architecture and make more sophisticated conversational workflows easier to build.

For enterprises, the combination of real-time interaction, telephony and backend delegation could support applications ranging from customer service and reservations to specialised business workflows.

However, the actual benefits will depend on factors such as integration design, reliability, latency, backend model selection and the requirements of individual applications.

The announcement provides OpenAI's evaluation results and selected customer examples, but independent comparisons across competing voice platforms were not provided.

The Bigger Shift Toward Conversational Agents

The release reflects a broader industry movement toward AI systems that do more than answer individual prompts.

Voice agents increasingly need to understand context, handle interruptions, use tools and continue working while interacting with users.

GPT-Live-1's architecture is built around that model of interaction.

The important change is therefore not simply that AI can speak more naturally. It is that the conversation itself can become the control layer for an AI agent that can delegate reasoning and actions to other systems.

For developers building voice-first applications, that creates a new design pattern: a real-time conversational model at the front, specialised reasoning and tools behind it, and telephony or application interfaces connecting the agent to the real world.

OpenAI's GPT-Live-1 release puts that architecture directly into its API and makes it available for developers building production voice experiences.