GPT-Live-1: OpenAI’s Full-Duplex Voice API Explained

OpenAI’s GPT-Live-1 is a developer-facing voice model built for conversations that feel less like a sequence of recorded turns and more like a live exchange. Available in the API from September 10, 2026, it costs $0.05 per minute for the front-end voice layer. Teams still pay separately for the reasoning model, tools, and other infrastructure they connect behind it. 35

What “full duplex” changes

Many voice agents use a cascade: speech-to-text converts a caller’s audio into text, a language model produces a text response, and text-to-speech turns that response back into audio. That approach can work well, but each stage is a handoff. Applications must coordinate timing, pauses, changing intent, and barge-ins across separate components. 3

GPT-Live-1 is positioned as a single voice model that reasons over incoming and outgoing audio. In practice, its full-duplex design means it can continue listening while generating speech. That is intended to let it respond to an interruption or a brief acknowledgement without treating every sound, silence, or nearby conversation as a new, mandatory turn. 3

The point is not simply faster audio generation. It is conversational control: recognizing when a person is still thinking, when they are cutting in, and when a short “yeah” is only a backchannel rather than a request for a new answer. OpenAI describes the model as its premier option for natural, expressive voice conversations with smooth interruption handling. 23

A voice layer paired with separate reasoning

GPT-Live-1 does not require the voice model to perform every piece of business logic itself. It can maintain the real-time conversation while delegating deeper reasoning, tool calls, or actions to a separately selected backend model and agent framework. 3

This separation gives developers a practical design choice:

• Use a faster, lower-cost backend for routine tasks such as scheduling or checking an order.
• Escalate difficult requests to a stronger model, such as GPT-6 Astra.
• Connect supported third-party backend models when that better fits an existing stack. 3

The result is a two-part system: GPT-Live-1 manages the live voice interaction, while the backend supplies reasoning, business data, and actions. The trade-off is that the advertised voice price is not the full cost of an agent; backend inference, tools, telephony, and related services can add to it. 35

API features and deployment options

OpenAI says developers can guide the model’s tone, pace, and conversational style through the system prompt. GPT-Live-1 also provides transcripts and response text, supports alphanumeric understanding and keyword biasing, and offers turn detection for products that still need explicit turn boundaries. 3

The model is intended for longer sessions and phone-based agents, including customer-support and reservation workflows. OpenAI says it can handle silence and background noise without narrating every internal step, a useful behavior for real phone calls and other less-controlled environments. 3

The supplied documentation does not establish a definitive supported-language list or detailed developer controls for specific accents and dialects. Those implementation details should be verified in the current API documentation before they are treated as deployment requirements.

Reported benchmark results

OpenAI reports that GPT-Live-1 improved by 30 percentage points over GPT-Realtime-2.1 on Full Duplex Bench, an evaluation focused on interactive voice behavior including turn-taking, pauses, interruptions, backchannels, and background speech. 3

When paired with GPT-6 Astra at medium reasoning effort, OpenAI also reports that GPT-Live-1 ranked first on Tau3, an end-to-end voice-agent benchmark. Its customer-service component covers spoken tasks in airline, retail, and telecom settings. 3

Some secondary reporting gives exact figures for turn-taking latency and Tau3 scores, but the primary-source material supplied here does not provide those precise values. The more defensible conclusion is that OpenAI reports substantial interactive-behavior gains and a top Tau3 result for the GPT-Live-1-plus-Astra configuration, not that every deployment will reproduce a specific benchmark number. 37

How early adopters are using it

OpenAI highlights several early voice-agent deployments:

• Yelp is using GPT-Live-1 in Yelp Host for inbound restaurant calls, including reservations and food orders. Yelp reports improved turn-taking and accuracy compared with its prior conventional voice architecture, along with better call-handling rates and more complete caller responses. 3
• Speak uses it in its language-learning tutor. OpenAI reports that learners can take longer to formulate answers and that interruptions fell by almost 80% compared with Speak’s earlier turn-based system. 34
• An unnamed healthcare company said moving away from a cascaded architecture reduced its codebase by 80% and removed 23,000 lines of code, supporting more natural real-time patient conversations from appointment booking to care navigation. This is a company-reported outcome, not an independently audited result. 3

These examples show where full duplex is most valuable: situations where a caller may hesitate, self-correct, acknowledge an answer mid-sentence, or interrupt an agent without wanting to restart the entire interaction.

Pricing and the role of OpenAI Presence

GPT-Live-1 is priced at $0.05 per minute for the front-end voice layer. Developers can pair it with their preferred backend model and agent harness, but those choices affect total operating cost because they are billed separately from the voice layer. 35

OpenAI also identifies OpenAI Presence as another way to build voice workflows using GPT-Live-1 for real-time voice interaction. The supplied material does not define Presence’s enterprise eligibility, commercial terms, hosting model, or access process, so those details should not be assumed from the API announcement alone. 3

For teams evaluating the model, the central question is therefore broader than the per-minute rate: how much value does smoother turn-taking create for the workflow, and what backend model and tool path delivers the necessary reliability at an acceptable total cost?