OpenAI Launches GPT-Live, a Voice Model That Listens and Talks at Once

OpenAI rolled out its new voice model, GPT-Live, on July 8, replacing the existing Advanced Voice Mode in ChatGPT worldwide. The biggest change is architectural: the model runs in full duplex, meaning it listens and speaks at the same time instead of waiting for a pause to start responding.

The previous voice mode worked on a turn-based basis: the user spoke, the model waited, then it replied. GPT-Live processes audio continuously, analyzing both the content of speech and its intonation, and decides many times per second whether to speak, listen, pause, interrupt, or reach for a tool.

How full duplex works

In practice, this means the model can interject a brief acknowledgment like mhmm or got it while the user is still talking, instead of waiting for the sentence to end. It can also go quiet when the other person needs a moment to think, which is meant to make the exchange feel like a natural conversation rather than a back-and-forth of messages.

For more complex questions that require web search or deeper reasoning, GPT-Live hands the task off in the background to GPT-5.5, and the result feeds back into the conversation once it's ready. Three reasoning levels are available, Instant, Medium, and High, letting the system balance response speed against quality.

The benchmark numbers

OpenAI published benchmark results in which participants rated conversations conducted through both voice modes. GPT-Live-1 beat Advanced Voice Mode in 75.7 percent of comparisons, while the lighter GPT-Live-1 mini won in 69.2 percent. In the Tau3 Voice Telecom test, which checks execution of phone-based tasks, the model at its highest reasoning setting achieved about 65 percent success with task completion times of around 385 seconds.

The biggest jump shows up in tasks that require knowledge and search. The score on GPQA, which tests PhD-level scientific reasoning, rose from 45.3 to 84.2 percent. On BrowseComp, which measures the ability to independently search for information online, performance jumped from 0.7 to 75.2 percent, underscoring just how limited the previous voice mode was in this area.

New features and availability

The model also gained nine reworked voices, visual cards that can display things like weather, stock quotes, or sports scores during a conversation, and live translation. For younger users, OpenAI introduced age-appropriate responses and parental controls, and safety mechanisms are meant to stay active for the entire duration of a voice conversation.

GPT-Live is rolling out for now to the mobile apps and the browser version of ChatGPT; API access is expected later, and the company is currently collecting sign-ups from developers interested in early access.

What it means for Polish users

For Polish companies building products on top of OpenAI's API, this signals a new category of voice assistants capable of holding a natural phone conversation or handling customer service without the typical pauses and rigid question-and-answer rhythm. For now, though, the feature is rolling out only to the consumer ChatGPT app, so commercial deployments will have to wait for the API to open up.

The GPT-Live launch is part of a string of OpenAI announcements from the first half of July, which also included the GPT-5.6 model family and the ChatGPT Work app, together painting a picture of a company rebuilding its product around three pillars: conversation, work, and coding.

Sources: Introducing GPT-Live (openai.com), ChatGPT can now listen and talk at the same time (the-decoder.com), OpenAI Launches GPT-Live-1 (mlq.ai)