GPT-Live-1

Overview

GPT-Live-1 is OpenAI's real-time voice model for natural, full-duplex conversations through the API, released on 2026-09-10. Its headline advance is that it is built for live speech-to-speech interaction rather than a traditional pipeline that separately combines speech recognition, text reasoning, and text-to-speech. The result is a model aimed at more fluid spoken exchanges, faster conversational turn-taking, and more consistent adherence to developer instructions in voice-first applications.

The model is especially notable for telephony-oriented experiences: support flows, appointment handling, call routing, sales assistance, coaching, and other scenarios where users expect to interrupt, clarify, and speak naturally. Full-duplex behavior is central: GPT-Live-1 is designed for conversations where listening and speaking can overlap, enabling barge-in handling and more natural back-and-forth than push-to-talk or turn-locked voice systems.

Key capabilities

GPT-Live-1 supports real-time audio interaction and speech-to-speech generation, allowing applications to stream user audio to the model and receive generated voice responses with low perceived latency. It emphasizes stronger instruction following, which is important for production agents that must follow scripts, respect escalation rules, avoid prohibited claims, or maintain a brand-specific support style.

Another major feature is custom voice support. This allows developers to create more distinctive voice experiences than generic assistant voices, such as a consistent service persona for a call center or a branded voice for an interactive product. In telephony contexts, the model's value is not only its voice quality but also its ability to combine spoken interaction with task-oriented behavior, such as gathering structured information, summarizing calls, or invoking backend tools when integrated by the developer.

Technical specifications and availability

• Provider: OpenAI
• Release date: 2026-09-10
• Primary modalities: real-time audio input and audio output, with speech-to-speech interaction and instruction-controlled behavior
• Typical interface: OpenAI API real-time voice workflows
• Context window: not specified in the provided model details; practical limits may depend on API session configuration and OpenAI's current deployment parameters
• Max output: not specified in the provided model details; output length is likely governed by streaming session settings, response limits, and voice interaction constraints
• Pricing: not specified in the provided model details; developers should consult OpenAI's current pricing page for audio, text, and real-time session costs
• Open-weight status: closed-weight proprietary model; no open weights indicated
• License/access: API access under OpenAI's applicable service terms rather than a self-hosted open-source license

Strengths and benefits

GPT-Live-1's main strength is reducing the friction of building voice agents. A single real-time model can simplify architectures that otherwise require separate ASR, LLM, dialog manager, and TTS components. Its emphasis on instruction following also makes it better suited to enterprise and customer-facing deployments where consistency matters. For end users, the benefit is more natural pacing, fewer awkward pauses, and the ability to interrupt or redirect the conversation.

Its telephony focus is also important. Phone audio is constrained, noisy, and interruption-heavy, so a model optimized for live calls can be more useful than a general multimodal model adapted after the fact.

Limitations and caveats

GPT-Live-1 is still a closed API model, so organizations cannot inspect weights, self-host it, or fully control runtime behavior. Real-time voice systems are sensitive to network latency, microphone quality, background noise, accents, and telephone bandwidth. The model may still misunderstand speech, produce incorrect information, or follow ambiguous instructions imperfectly. Cost can also be higher than text-only models because audio streaming and long calls consume more resources.

Custom voices raise governance questions around consent, impersonation, disclosure, and brand safety. Deployments in regulated environments should include logging, human escalation, monitoring, and clear user disclosure that they are interacting with an AI system.

Comparison

Compared with earlier OpenAI real-time voice offerings, GPT-Live-1 appears more directly optimized for production voice agents, with stronger instruction following and custom voice experiences as explicit priorities. Compared with alternatives such as Google Gemini Live API, ElevenLabs conversational voice systems, or modular Twilio-plus-ASR/TTS stacks, GPT-Live-1's appeal is the integration of conversational intelligence and real-time voice generation in one OpenAI model. However, specialized voice platforms may still compete strongly on voice design tooling, telephony integrations, or deployment controls.