Constructing conversational AI voice systems has traditionally required orchestrating multiple interconnected components. Speech input must be converted to text, processed by a language model to generate a response, and then synthesized back into audio—a pipeline that often results in stilted, robotic interactions. OpenAI is attempting to streamline this architecture with a fundamentally different approach.

The company unveiled GPT-Live-1 through its API on Wednesday, making the full-duplex voice technology underlying ChatGPT's voice mode available to external developers for the first time. Rather than requiring developers to manage every component of the voice pipeline independently, the system designates a single model for conversational interaction while offloading intensive computational tasks to separate backend systems.

Full-duplex voice delegation

GPT-Live-1 serves as the primary conversational interface. Its native full-duplex capability enables it to process ongoing dialogue in real time, including handling interruptions and overlapping speech, without requiring developers to coordinate multiple distinct systems. The voice model need not shoulder the entire processing burden independently.

When a query demands extended processing time or greater computational resources, GPT-Live-1 can transfer the request to another model operating in the background. This could be GPT-6 Astra, a more compact model such as Luna, or a third-party offering.

Waiting for a more powerful model to complete processing can render voice interactions uncomfortably awkward. A complex question might leave the user sitting through silence while the backend model deliberates. GPT-Live-1 maintains conversational continuity instead—filling gaps, acknowledging the speaker—before integrating the answer once the backend completes its work.

According to OpenAI, GPT-Live-1 achieves performance 30 percentage points higher than GPT-Realtime-2.1 on Full Duplex Bench. When combined with GPT-6 Astra configured for medium reasoning, it also ranks first on the 𝜏³-benchmark.

What the handoff looks like

OpenAI implements delegation through an event-driven architecture. The voice session generates a delegation_id, transmits context to the backend system managing the heavier computation, and receives the result via an event designated session.commentary.append. The voice model incorporates that result into the ongoing exchange rather than reading a complete text block aloud. Developers retain visibility into what the model perceives and produces and can govern when it speaks—they simply avoid constructing the entire conversation from separate components. OpenAI's API documentation provides comprehensive guidance on this pattern, complete with a functional example using the Codex SDK.

Early customers cut code

One early adopter eliminated 23,000 lines of code upon migrating to GPT-Live-1.

Tony Stoyanov, co-founder and CTO at EliseAI, a healthcare organization piloting the API, reported that the transition reduced his team's codebase by 80%. This freed his engineers to concentrate on enhancing the patient experience—streamlining appointment booking and care navigation.

Language-learning platform Speak observed tangible improvements in actual conversations. During initial testing of its Live Tutor Lessons feature, GPT-Live-1 demonstrated nearly 80% fewer interruptions when users paused momentarily to formulate responses. For language learners, those additional seconds frequently determine whether they can articulate an answer or get cut off by the AI.

Yelp has already integrated GPT-Live-1 into Yelp Host and Hatch. CTO Alex Levy noted that the platform is handling more calls through AI successfully, with callers using fuller, more natural speech patterns—an indication, according to Levy, that the overall phone experience feels distinctly different. A demonstration accompanying the announcement illustrates this effect clearly, showing a restaurant reservation system maintaining continuity despite background noise and overlapping speakers.

Pricing the voice layer

GPT-Live-1 carries a cost of $0.05 per minute, equivalent to approximately $3 hourly. Beyond this, developers incur charges for whatever backend systems they select. Should GPT-Live-1 route a request to GPT-6 Astra, the developer pays for that invocation as well. Each delegation to a reasoning model accelerates cost accumulation.

The more often an agent reaches for a reasoning model, the faster the bill climbs.

OpenAI has been reducing API pricing amid intensifying competition from Anthropic, Google, and Chinese research labs, yet advanced reasoning capabilities remain a paid service.

The advantage lies in enabling developers to make strategic decisions about expenditure allocation. Straightforward tasks, such as scheduling appointments, could utilize Luna. More demanding queries requiring multi-step reasoning or tool integration could leverage Astra. OpenAI has previously demonstrated how Astra's configurable reasoning parameters allow developers to adjust expense per request, and GPT-Live-1 furnishes a framework for applying this same optimization logic to voice interactions.

Platform control tradeoffs

Under the traditional cascaded architecture, teams could select distinct providers for each voice stack component and substitute elements as needed. GPT-Live-1 assumes greater responsibility for the conversation, which simultaneously means delegating more of it to OpenAI.

The bet is that developers will give up some of that control if it means voice agents can finally keep up with the people talking to them.

Source: The New Stack