OpenAI is rolling out its new ChatGPT voice model, GPT-Live, to all users in stages.

Tech site TechRadar reported on July 9 that the key to the update is a two-way conversation feature that lets users and ChatGPT speak and listen at the same time.

The model is provided in different versions for free and paid users. Free users receive GPT-Live-1 Mini, while paid users receive GPT-Live-1. Users can tell it is applied if a “Live” label appears at the top of the mobile voice chat screen or behind the ChatGPT selection menu on the web. They can also tap the label to switch back to the previous voice model.

Users can also check whether it is applied in the settings menu. On mobile, they can open the voice menu from settings or their profile behind the menu button and check the model. On the web, they can check in the order of settings, voice and model from the profile at the bottom left.

The biggest change is the flow of conversation. Users can keep talking even if ChatGPT starts responding before they finish, and the service is designed to follow that flow. If users cut off an answer midway and ask a new question, it adjusts its response again. It is closer to how people talk than the previous voice mode.

A real-time translation feature has also been placed front and center. When a user speaks a sentence, ChatGPT can immediately translate it and speak it in a foreign language.

Users can also adjust how voice conversations respond. They can choose an intelligence level along with the voice from an adjustment icon at the top right of the screen. The options are Instant, Medium and High. OpenAI recommended using Instant for quick answers, High for detailed answers and Medium for balance. The voice types are the same as before.

Visual cards have also been added to voice mode. As a result, questions such as sports scores, weather forecasts and places found on maps display graphic information on the screen. It expands responses beyond audio alone by combining them with the screen.

Early user reactions have generally leaned toward the features being stable. Some also said it feels more like talking to a real person, including hesitations and varied speaking styles.

Some issues were also identified during the initial rollout. Ati Eleti (애티 엘레티) of OpenAI said in a related Q&A on Reddit that there is an issue in which ChatGPT memory does not work in some cases in voice mode, and the company is actively investigating it. Problems with pronouncing foreign languages with an English-style accent have also been reported, and Eleti said that part would also improve over time.

The update appears aimed at raising ChatGPT’s voice feature from a simple read-aloud assistant function to a real-time interactive tool. In particular, it broadened the scope of voice interfaces by bundling overlapping voice conversations, instant interruption, real-time translation and visual responses at once. Memory integration and stabilising multilingual pronunciation appear to remain tasks.