What Full-Duplex Voice AI Changes for a Business Phone Line

OpenAI opened its GPT-Live-1 speech model to developers. It listens and talks at the same time, costs $0.05 per minute, and still fails most of a banking support benchmark. Here is the honest read.

OpenAI has released GPT-Live-1 as an API, so developers can build voice applications on the model that already runs inside ChatGPT. As reported by The Decoder, the headline feature is full-duplex speech: the model listens and speaks at the same time, instead of waiting for the caller to finish. If your business takes bookings, order status calls or after-hours enquiries by phone, this is the part of the AI news cycle that actually touches your operations.

What actually changed

Until now, voice assistants worked in turns. You speak, it waits, it answers. Full-duplex removes that wall: the caller can interrupt, add a detail mid-sentence, or say "no, the other Tuesday" without the system losing the thread.

OpenAI's own benchmark numbers show the gap against its previous model, GPT-Realtime-2.1. Full-duplex interactivity scores 80.1 percent versus 45.4 percent. Turn-taking latency falls to 0.8 seconds from 1.4 seconds. Tool-calling accuracy — the model's ability to correctly trigger an action like a lookup or a booking — rises to 87 percent from 60 percent.

A few other practical details. Developers can pair GPT-Live-1 with different backend models, choosing how much reasoning, speed and cost each use case gets. The model ships with twelve new voices covering various accents, dialects and languages. It returns speech-recognition transcripts and response text by default, so every call comes with a written record. Yelp is already using it for phone reservations and its CTO, Alex Levy, reports better call handling.

The price is $0.05 per minute. That is not a rounding error at volume.

The number nobody is quoting

On a banking voice support benchmark, GPT-Live-1 passes 32 percent of cases. The previous model passed 12.4 percent. That is a large relative improvement and a bad absolute score. In a hard, regulated support scenario, this model still fails roughly two out of three times.

Read the two facts together and you get the right mental model. Voice AI has become genuinely good at the mechanics of conversation — timing, interruptions, sounding natural. It has not become reliable at complex, consequential problem solving. The gap between "feels human" and "gets it right" is where businesses will get burned this year.

Why it matters for a small business

Most small-business phone traffic is not complex. It is opening hours, is my order ready, can I move my appointment, do you have this in stock. These are exactly the calls that fit inside a 32 percent-on-hard-problems model, because they are not hard problems. They are lookups plus a booking action, and tool-calling accuracy at 87 percent is the relevant number there.

The cost is manageable if you scope it. At $0.05 per minute, a three-minute call costs 15 cents. Five hundred such calls a month cost about $75 in model fees, before development and the rest of your stack. Compare that to the calls you currently miss after 18:00, and the arithmetic is straightforward for some businesses and clearly negative for others. Do the calculation with your own call log, not with a vendor's.

The twelve voices with different accents and languages matter in markets where customers switch between languages mid-call. The built-in transcripts matter more than they sound: every call becomes searchable text you can audit, feed into a CRM, or use to find out what customers actually ask.

What to do about it

  • Pull one month of call records first. Count how many calls are simple lookups, how many are after-hours, and how many go unanswered. Without those three numbers you cannot judge whether a pilot pays for itself.
  • Pick one narrow job. Appointment booking or order status, not "handle the phone." Narrow scope is what keeps you on the right side of that benchmark gap.
  • Define the handoff before you define the greeting. Decide exactly when the assistant stops and transfers to a person: any complaint, any refund, any repeated question. Write those rules down.
  • Require the assistant to be able to act, not just talk. It needs real access to your booking calendar or order database. Tool-calling accuracy is useless if there are no tools connected.
  • Use the transcripts from week one. Read them. They will show you where the assistant guessed, where it annoyed people, and which questions you should answer on your website instead.

Full-duplex is a real improvement in how a machine holds a conversation, and the API pricing puts it within reach of a small operation with steady call volume. It is not a replacement for the person who knows your business. Treat it as a filter on the front of your phone line that handles the repetitive third of your calls, keeps a written record, and hands the rest to a human quickly.

SourceWritten from reporting by The Decoder. Read the original: OpenAI's GPT-Live-1 API lets developers build apps that talk and listen at the same time

More articlesAll articles

Meta's Muse Shops and Negotiates on WhatsApp for Buyers

Meta launched Muse, an AI agent controlled through WhatsApp that browses, fills forms, negotiates and pays via Stripe's Link. Here is how to prepare your catalog, prices and support scripts.

Read

How Much Does a Website Cost for a Small Business?

A website has two price tags: the one-time build and the monthly cost of keeping it alive. Here is what sits inside each, where a builder is enough, and when custom code actually pays for itself.

Read