GPT-Live-1 reaches the API and can listen while it speaks
Original publication: · Sergey Zhadov
This is a translated archive article. The date and pricing discussion refer to the original publication, not a new announcement today.
OpenAI has released GPT-Live-1 in the API. The model can listen and speak at the same time, cope better with interruptions, pauses and background noise, and keep a conversation moving without the stiff stop-start rhythm of a basic phone bot.
Why a conversation without turn taking matters
A traditional voice bot often chains speech recognition, a text model and speech synthesis. Every handoff adds delay and another chance to lose timing or context. GPT-Live-1 processes incoming and outgoing audio together, so it can react when a caller interrupts, hesitates or changes direction mid-sentence.
That does not guarantee a flawless call. It does give developers a sturdier base: the model is designed to handle silence, short acknowledgements and nearby speech without blurting nonsense over the user. For support, reservations and phone forms, that is a useful improvement rather than decorative AI fluff.
The voice stays fast while another model does the heavy work
The voice layer can keep the exchange alive while a backend model checks an order, reasons through a complaint or calls a tool. Developers choose the model and harness behind it, which means simple status requests do not need the same expensive route as a complicated customer problem.
This is also where teams can screw up. Send every tiny request to the priciest reasoning model and the natural voice arrives with a brutal invoice. A sensible build classifies tasks, limits tool calls and escalates only the cases that genuinely need more brainpower.
The early numbers look strong but need real testing
OpenAI reports a 30 percentage point improvement over GPT-Realtime-2.1 on Full Duplex Bench. One early customer also reported almost 80 percent fewer interruptions while language learners paused to think. Those figures are promising, but they come from the developer and its partners.
Run the model against your own microphones, accents, noisy rooms, weak connections and impatient callers. A tidy benchmark will not expose every ugly edge case. A few hundred real conversations usually reveal where turn detection, transcripts or tool routing still fall apart.
What it costs and what to measure before launch
The front-end voice layer is available in the API for $0.05 per minute. Reasoning models, tools, hosting and phone service add their own charges. Multiplying call minutes by one headline price and calling it a budget is cheerful nonsense.
Before production, measure response latency, successful interruption handling, cost per resolved contact and transfers to a human. The takeaway is simple: GPT-Live-1 can make voice agents feel less wooden, but one model does not fix a bad workflow. Without routing, limits and review, the mess merely learns to speak nicely.
Translation reviewed:
