On September 10, OpenAI brought GPT-Live-1 into its API, a voice model built to fix one of the most annoying traits of phone bots: constantly talking over the caller. Instead of processing listening, reasoning and speaking one after another, GPT-Live-1 handles incoming and outgoing audio at the same time in a single model, what OpenAI calls “full-duplex,” borrowing the term from phone lines that carry signals in both directions at once. Early customers such as review platform Yelp and language-learning app Speak report noticeably more natural conversations.
Key takeaways
- GPT-Live-1 processes speech in a single model instead of the previous chain of speech recognition, language model and speech synthesis, letting it listen and speak at the same time.
- On OpenAI’s own Full Duplex benchmark, the score jumps from 45.4 to 80.1 percent compared with predecessor GPT-Realtime-2.1, while response latency drops from 1.41 to 0.80 seconds.
- Language-learning app Speak reports nearly 80 percent fewer interruptions during learners’ thinking pauses.
- The API costs $0.05 per minute for the voice layer alone; more complex tasks are handed off to paired models such as Astra or Codex, which are billed separately.
- The target use cases are phone agents, support, scheduling and tutoring, anywhere natural interruption handling and fast responses shape how a conversation feels.
What “full-duplex” actually changes
Most voice assistants today work in half-duplex fashion: they listen, wait for a pause, process what was said, and only then respond, much like a walkie-talkie. That is the source of familiar phone-bot annoyances: cut-off sentences, awkward pauses, replies that collide with what the caller was about to add. GPT-Live-1 processes audio continuously in both directions, so it can tell whether a speaker is just searching for words or genuinely done talking. In OpenAI’s own tests, the clearest gain shows up in turn-taking: response time dropped from 1.41 to 0.80 seconds, and in a banking test scenario, accuracy rose from 12.4 to 32 percent. The model handles only the voice layer, including background noise suppression, context retention across longer sessions, and recognition of digits and letters.
Who is already using it
Yelp uses GPT-Live-1 for phone reservations and orders. CTO Alex Levy describes the effect as callers now speaking “fuller, more natural sentences,” a sign that the experience on the other end of the line feels genuinely different. At language-learning app Speak, which has to deliberately give users pauses to think out loud, interruptions dropped by almost 80 percent compared with the previous turn-based system, according to co-founder Andrew Hsu. Customer service provider Fin describes a shift away from a “stop-start rhythm” toward the natural flow of an actual phone call. All three examples point to the same soft spot in existing voice bots: it is not the cleverness of the answer that shapes the impression, but the timing.
A building block, not an all-in-one agent
OpenAI deliberately positions GPT-Live-1 as a pure voice layer rather than a standalone agent. For scheduling or routine requests, it can be paired with the leaner Luna model; for more complex customer issues, with Astra; and for coding tasks that come up mid-conversation, even with Codex — each connection billed separately, so the headline $0.05 per minute covers only part of the actual running cost. That mirrors the architecture behind OpenAI’s Agents API, opened just days earlier for text-based tasks: OpenAI supplies the infrastructure, and developers build the actual application on top. For companies already experimenting with voice bots, the ability to mix and match models is likely the more interesting story than the voice quality itself, echoing the gap between speed and polish that has already shown up in retail AI advisory tools.
What it means for users
For consumers, the difference will stay invisible until companies actually build the technology into their support lines and booking systems, a process that typically takes months. At five cents a minute, GPT-Live-1 costs well above plain text AI, making it most attractive for use cases where a better conversation directly translates into revenue or customer satisfaction, such as reservations, banking or language tutoring. Whether the benchmark gains actually translate into fewer frustrated callers in everyday use remains to be seen; the customer testimonials so far are, unsurprisingly, hand-picked from companies working closely with OpenAI. It is also worth noting that OpenAI keeps the pricing deliberately modular: anyone who only needs the voice layer pays little, while anyone who wants full task-handling capability pays again for every additional model they connect — a building-block approach that gives developers flexibility but makes total costs harder for businesses to predict than a simple flat fee per call.