OpenAI's newest voice AI just moved out of the chat app and into the phone system. On September 10, 2026, OpenAI made GPT-Live-1 available through its API, the version of the model other companies can build their own products on top of. It is the same voice technology that started rolling out inside ChatGPT two months earlier, and it is built to hold a real back-and-forth conversation instead of a stiff, one-question-at-a-time exchange.
OpenAI's newest voice model is called GPT-Live-1. It launched inside ChatGPT in July 2026, then became available to developers in OpenAI's API on September 10, 2026. It can listen and speak at the same time, which OpenAI says makes interruptions and natural back-and-forth possible. Your business cannot use it directly today; software companies build phone products with it.
What OpenAI actually released
OpenAI introduced GPT-Live-1 and a smaller GPT-Live-1 mini inside ChatGPT on July 8, 2026, replacing the app's earlier Advanced Voice Mode, as reported by TechCrunch. The older system chained together separate steps: transcribe the caller's speech to text, run it through a language model, then convert the reply back to speech. GPT-Live-1 replaces that chain with what OpenAI calls a full-duplex model, one system that can speak and listen at once.
Two months later, OpenAI opened a version of that model to developers. Its own API changelog records the change plainly: "GPT-Live 1 is now generally available in the API. Build full-duplex voice conversations that can continue while a backend model or agent handles reasoning and tools." That is the release this article is about, because it is the point where GPT-Live-1 stops being something you only talk to inside ChatGPT and starts being a part other companies, including phone-answering products, can build with.
For context, GPT-Live-1 is not OpenAI's first realtime voice model. It followed gpt-realtime-2.1, which OpenAI released on July 6, 2026, and described in its own developer forum as cutting "p95 latency by at least 25% across Realtime voice models through improved caching." GPT-Live-1 is the newer, full-duplex generation.
How it actually behaves on a call
The headline feature is full duplex: the model listens while it is talking, the way two people do on a real phone call, instead of waiting its turn like older voice assistants. In OpenAI's own description, GPT-Live-1 "listens and speaks simultaneously, handles pauses, interruptions, and backchannels" and can make sounds like "mhmm" or "yeah" while the caller is still speaking, so a caller does not feel like they are being cut off or ignored.
A caller no longer has to wait for a beep. The model is designed to listen while it talks, the same way a person does.
For businesses that route calls through software, the model also supports telephony and SIP alongside the web-based connections developers already use, according to OpenAI's API announcement, and ships with twelve built-in voices. It can hand reasoning and tool use, like checking a calendar or looking up a price, off to a separate backend model or agent while the conversation keeps running.
OpenAI also published its own comparison numbers against the previous generation. Paired with GPT-6 Astra, GPT-Live-1 completed 83.6% of a benchmark called Tau3 on the first attempt, compared with 45.7% for GPT-Realtime-2.1, and OpenAI reported turn-taking latency of about 0.8 seconds versus 1.4 seconds for the older model. Those are OpenAI's own published figures, not an independent test, so treat them as a company's claim about its own product rather than proof of how it performs on your specific calls.
Who can actually use it
There are two very different audiences here, and it matters which one you are.
ChatGPT's voice mode now uses GPT-Live-1 mini by default, replacing the earlier Advanced Voice Mode, as reported by TechCrunch. That is the version a person talks to directly in the ChatGPT app. It is not something a business can point at its own phone line.
In the API, GPT-Live-1 is priced for developers: OpenAI's changelog and model documentation put voice sessions at $0.05 per minute, billed per second, with any backend model or tool use billed separately. That is the version software companies use to build a product, the same way they would use any other OpenAI model. A local business does not get a GPT-Live-1 account any more than it gets a direct account with the engine inside a piece of software it buys. Someone still has to build the phone-answering product on top of it.
What this means for your phone
Two things change here for a plumber, HVAC company, electrician, or roofer, even without touching this model directly.
First, expectations move. Every person who talks to a natural, interrupt-friendly voice assistant in their own pocket gets a little less patient with an automated system that talks over them, mishears a phone number, or forces them through a rigid menu. That expectation carries over to your business's phone line, whether it is answered by a person, a script, or software.
Second, the building blocks got cheaper and more available. Because GPT-Live-1 is now sold through the API with telephony support built in, expect more vendors, including AI receptionist products, to build call-answering tools on top of it or a model like it over the coming months. Onvertz's own product, Loop, is one example of that category: per its product page, it answers calls, books appointments into a calendar, transfers urgent calls, and texts follow-ups, logging each call as a transcript and summary. The point is simply that this kind of tool exists and is getting easier for software companies to build well.
Which model runs underneath a phone product matters less than what the caller experiences: did the call get answered, was the price explained correctly, and did the job get booked. That is true no matter which AI voice OpenAI or anyone else ships next.
What to watch for before you trust it on a live call
A smoother-sounding conversation is not the same as a more accurate one. Full duplex and lower turn-taking latency describe how natural a call feels, not whether the AI quoted the right price or booked the right time slot. Any phone system, OpenAI's or anyone else's, still needs clear rules about what it is allowed to promise, quote, or confirm on your behalf.
The advertised price is not the whole cost. $0.05 per minute in OpenAI's API covers the voice layer only; a working phone agent also needs a backend model to reason and use tools, a telephony connection, and someone to build and maintain the integration. Sticker price is not the total bill.
More natural AI voices cut both ways. If callers get more comfortable talking to AI, some of the calls reaching your team will be scammers or robocallers using the same kind of natural-sounding voice technology. Train staff to verify who they are actually speaking to before sharing account details or taking payment over the phone, the same caution that already applies to any unexpected caller.
OpenAI's voice models are moving fast: two model generations and a public API release in about ten weeks this summer. Businesses do not need to chase every release. They do need a phone system, whatever runs it, that answers reliably, quotes honestly, and hands off cleanly when a human needs to step in.