OpenAI has started deployment GPT-Live, a new generation of voice models that power ChatGPT Voice. During a conversation, GPT-Live can forward a question to its latest frontier model, such as GPT-5.5, for further reasoning or web searching.
What was launched
GPT-Live-1 becomes the default model powering ChatGPT Voice for Go, Plus, and Pro users, and GPT-Live-1 mini becomes the default model for free users.
OpenAI describes the models as full-duplex, meaning they can capture audio and produce speech at the same time, interrupt less, and wait when a user pauses.
Search and visual responses enter the conversation
When a spoken question requires deeper reasoning or current information, GPT-Live can pass it to a border model, GPT-5.5 at launch, and bring the answer back into the conversation.
ChatGPT Voice can also display visual maps for topics like weather, stocks and sports, and OpenAI said Voice continues to support search, memory, images and file downloads. Users can set a level of reasoning, with Instant for quick answers and Medium or High for more complex questions. OpenAI said Instant Replies and Mini Model run on GPT-5.5 Instant, while Medium and High use GPT-5.5 Thinking.
How OpenAI measured it
In direct evaluations of OpenAI, GPT-Live-1 and GPT-Live-1 mini were preferred over its advanced voice mode in conversations lasting five to ten minutes. OpenAI said the assessments looked at overall preferences, turn-taking, interruptions, conversation flow and naturalness of interactions. The company said more than 150 million people talk to ChatGPT every week using features like voice and dictation.
What he hasn’t done yet
At launch, GPT-Live does not support voice with video or screen sharing in ChatGPT. OpenAI said it is working to add these features and that its old standard and advanced voice modes remain available where video and screen sharing are supported.
Why it matters
Now, when you ask a question out loud, you can see an answer directly in the voice stream, sometimes accompanied by a visual map on the screen. This provides ChatGPT with another way to present a response without visiting a source site. Since the reasoning and search process happens behind the scenes on GPT-5.5, what appears there depends on how the model retrieves information and cites its sources.
Looking to the future
OpenAI said the rollout is starting globally and video and screen sharing are not in this release but are being worked on.
The OpenAI article does not specify how GPT-Live handles citations when responding to a spoken question from a GPT-5.5 web search. ChatGPT text responses display source links next to the response. Whether an oral response names its sources, displays them on screen or leaves them aside is the detail to watch out for. This is what decides whether a search in a voice chat can still send a reader to your site.
Featured image: OpenAI
