OpenAI’s GPT-Live-1 API: Full-Duplex Voice, Backend Delegation, and $0.05/Minute Pricing

OpenAI announced the general availability of GPT-Live-1 in its API on September 10, 2026, not September 11. The release gives developers a voice layer for building agents that can listen and speak at the same time, instead of treating a conversation as a sequence of rigid, completed turns. 116

What “full-duplex” changes

Many voice-agent systems are built as a chain: speech-to-text converts the caller’s audio into text, a language model produces a response, and text-to-speech turns that response back into audio. GPT-Live-1 is positioned as a unified voice layer for the real-time conversational part of that interaction. 1

The practical distinction is turn-taking. OpenAI says GPT-Live-1 can listen and speak simultaneously, follow pauses and shifts in pace, and decide whether it should answer or continue listening. That design is intended to make interruptions and brief acknowledgements—often called backchannels—feel less like errors and more like normal conversation. 1915

This does not mean the voice model must perform every task itself. A GPT-Live-1 conversation can continue while a backend model or agent performs deeper reasoning or uses tools. OpenAI supports Responses delegation with an OpenAI model and client delegation for a developer’s own backend, enabling teams to choose the reasoning system behind the voice experience. 116

How the API is structured

GPT-Live-1 acts as the front-end conversational voice layer. The backend can be responsible for tasks such as more involved reasoning, retrieval, workflow execution, or tool calls. In OpenAI’s own product notes, GPT-6 Astra is identified as an option for harder search and reasoning tasks; API developers can also connect their own backend through client delegation. 416

For builders, that separation creates a trade-off:

• GPT-Live-1 manages the live interaction: listening, speaking, and conversational timing.
• A backend model or agent can handle complex decisions and tools.
• The overall agent cost and behavior depend on both layers, not on the voice model alone. 1618

OpenAI also highlighted stronger instruction following, custom voices, and telephony support for the API release. 13

Pricing: $0.05 per minute for voice, plus backend usage

GPT-Live-1 voice sessions cost $0.05 per minute. Billing is calculated per second, with no rounding up to the next full minute. 1618

That is the price for the voice session—not necessarily the complete cost of a voice agent. Backend model usage and tool usage are billed separately, so a production budget needs to account for both the duration of the conversation and the work delegated behind it. 1618

Performance claims and what the supplied evidence supports

OpenAI reported that GPT-Live-1 improved by 30 percentage points on its Full Duplex Bench compared with GPT-Realtime-2.1, and that GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranked first on Tau3. 1

However, the supplied source material does not independently substantiate the specific figures in the original question for turn-taking latency or the exact Tau3 pass-rate comparison. Those precise numbers should not be treated as verified here.

What early users reported

OpenAI described several early production deployments as evidence for the value of more natural turn-taking:

• Yelp reported better turn-taking and accuracy after integrating GPT-Live-1 into Yelp Host and Hatch. For reservation and food-order calls, it said callers used fuller, more natural utterances and call handling improved. 1
• Speak reported that learners could take more time to think before a tutor responded, reducing interruptions by almost 80% compared with its previous turn-based systems. 1
• A healthcare company said moving away from a cascaded implementation simplified its codebase by 80% and removed 23,000 lines of code, while supporting more natural real-time patient conversations. 1

These are company-reported outcomes rather than independent evaluations, so they are best read as deployment examples—not universal performance guarantees.

What remains unconfirmed in the supplied material

The available evidence does not substantiate claims about enterprise access through a managed product called “Presence,” nor does it establish a specific expansion in accents, dialects, or languages for this API launch. Those details should be verified directly in current OpenAI documentation before being used in procurement or implementation decisions.

Bottom line

GPT-Live-1’s main change is architectural: it brings a full-duplex conversational voice layer to the API, allowing an agent to keep listening and speaking naturally while a separate system performs slower reasoning or calls tools. At $0.05 per minute for the voice layer, billed by the second, it also makes the cost of live voice time straightforward—but the complete cost still depends on the selected backend model and tools. 11618