OpenAI Opens GPT-Live-1 Voice Model That Talks and Listens at Once

OpenAI has released GPT-Live-1 to developers through its API, TechRepublic reported on September 15, 2026. The model is a full-duplex voice system: it can listen and speak at the same time, so a voice agent built on it can handle interruptions, pauses and changes of direction instead of waiting for a speaker to finish before answering.

Most voice assistants still follow a listen-then-answer pattern. A caller talks, the system waits for silence, and only then does it respond. GPT-Live-1 is designed to break that rhythm. According to TechRepublic, the model can keep a conversation going while connected models and backend tools handle the underlying work, then report the result back to the person on the line.

The numbers OpenAI reported

OpenAI reported a turn-taking latency of 0.798 seconds on its Full Duplex Bench and 87% tool-calling accuracy. Both figures come from the company itself, and Full Duplex Bench is OpenAI’s own test, so they are best read as the vendor’s claims rather than independent results.

Pricing is set at $0.05 per minute through the API platform. TechRepublic noted that backend reasoning models, tool calls and other components used alongside GPT-Live-1 can be billed separately, so what an application actually costs depends on how it is built.

Where American businesses might use it

OpenAI is pitching GPT-Live-1 as a building block for voice agents that do more than talk. TechRepublic described use cases across customer support, sales and reservations: a sales agent that keeps speaking with a prospect while it retrieves product information, or a support agent that stays on the call while backend systems work on a problem and then tells the customer where things stand.

For US companies that run large call centers, that shifts voice AI from a stand-alone feature toward an interface for broader AI agents. It also raises practical questions the source flagged: how callers are told they are speaking with an AI, how identity is checked, how errors are handled, and when a human should take over.

The fraud question

TechRepublic also pointed to a plausible misuse. A system that listens continuously and adapts its answers in real time gives scammers something prerecorded robocalls lacked: situational awareness during a live call. The outlet was careful to say GPT-Live-1 has not been shown to power such scams. Its point was that the barrier to building convincing automated voice interactions is falling, and that businesses and consumers may need stronger ways to verify who is on the other end of a call when a familiar voice is no longer proof.

What to watch

The open questions are how developers price and build on the model, whether independent testing matches OpenAI’s reported latency and tool-calling figures, and how companies handle disclosure and authentication as voice agents start taking real actions on a caller’s behalf.