What OpenAI actually shipped

OpenAI released GPT-Live on 8 July 2026 in two versions, GPT-Live-1 and a lighter GPT-Live-1 mini. The design is full-duplex, meaning the model listens and speaks at the same time and makes an interaction decision many times a second, so it can drop in a short acknowledgement, hold a quick back-and-forth, or simply stay quiet while you think. For anything that needs real reasoning or a web search, it hands the work to a frontier model in the background, GPT-5.5 at launch, and returns the answer into the conversation. It is rolling out now to every ChatGPT user across iOS, Android and the web, with GPT-Live-1 becoming the default voice for Go, Plus and Pro subscribers and the mini version the default for free users. In OpenAI's own testing, people strongly preferred it to the previous Advanced Voice Mode.

The catch is who can build on it. During this launch the models power the consumer ChatGPT voice experience, while developer access sits behind a signup form with no stated availability date. Video and screen sharing are not supported yet, and full multilingual parity is missing, with non-English demonstrations sounding noticeably off. So the most natural-sounding voice OpenAI has ever shipped arrived as a consumer feature first, and the endpoint a company would need to put it inside its own product is still a waitlist.

The best voice and the buildable voice have split

A week before GPT-Live, on 1 July, xAI shipped the Grok Voice Agent Builder, and the contrast is the story. It is a no-code surface plus an API that lets a team stand up a production voice agent in a couple of minutes, with telephony, knowledge retrieval, tools, guardrails and observability included. It is priced and available today at about 0.05 dollars, close to 4 pence, per minute of audio with the voices included, plus roughly 0.01 dollars a minute if you attach a provisioned phone number, and server-side searches billed separately. So there are now two different answers to the question of the best voice AI: the best experience, which is OpenAI's and is consumer-only for now, and the best one you can actually deploy, which ships with an API and a per-minute price you can put in a model.

The operator lesson is to stop confusing the demo with the deployable. A voice you cannot buy at scale, whose terms can change before general availability, is a signal about where the market is heading, not a dependency you can ship a customer line on. When you are choosing what a customer will hear on your support number this quarter, the shortlist is the set of models that come with a published API and a price, and right now the most complete of those is not the one that got the headlines.

What a European operator should do about it

A metered per-minute price is the useful part, because it turns voice into a cost line you can forecast. At about 0.05 dollars a minute, a four-minute support call is roughly 0.20 dollars, close to 16 pence, of voice before any tool calls, and that is a number you can set against the cost of a human minute and against your call volume. Treat it as a unit cost, not a novelty: model the monthly bill at your real call minutes before you commit, and remember that server-side searches and a phone number add to it. The point is that a priced API lets you do this arithmetic, and a waitlisted consumer feature does not.

Two cautions belong in the plan. First, the multilingual gap is not a footnote for a European business: the English demo does not guarantee the quality your customers will hear in their own language, so validate the agent in that language, on your own scripts, before it goes anywhere near a customer. Second, expect the full-duplex bar OpenAI just set to become the baseline customers assume within months, which means a turn-based phone menu will start to feel dated faster than you planned. Build on a model you can actually buy, price it at your real volume, test it in your market's language, and watch OpenAI's general release so you can move when the better voice becomes something you can deploy rather than only something you can show.