OpenAI's GPT-Live: A voice-enabled ChatGPT that can listen and respond.

      OpenAI wants users to engage with ChatGPT through conversation rather than typing. On July 8, it introduced GPT-Live, a new generation of voice models, which the company claims makes interactions with AI feel significantly more like a real dialogue. Two variants, GPT-Live-1 and a smaller version called GPT-Live-1 mini, are now available to ChatGPT users globally.

      The primary change lies in the architecture. GPT-Live operates in “full-duplex” mode, allowing it to listen and speak simultaneously. This capability enables it to interject with affirmations like “mhmm” or “got it,” engage in quick exchanges, or pause while you formulate your thoughts. Previous voice modes had to wait for users to finish speaking, often resulting in interruptions at inappropriate times.

      The second major advancement is in its delegation capabilities. When a question requires an online search or complex reasoning, GPT-Live refers it to a more powerful model operating in the background. At launch, that model is GPT-5.5. While waiting for the answer, GPT-Live continues the conversation and integrates the response seamlessly.

      “This mirrors how humans communicate with each other,” said Atty Eleti, the voice product lead for ChatGPT, to reporters as reported by CNET. “We maintain the flow of conversation while thinking in the background.”

      One notable demonstration featured the model performing real-time, simultaneous translation, providing a running translation as a presenter spoke. TechRadar, which witnessed the feature, described it as genuinely valuable. The model also responds to a wake word, with OpenAI staff using “Hey Chat” during the demos.

      OpenAI describes GPT-Live as its most advanced voice model to date. Users can select a reasoning level: Instant, Medium, or High. The voice can now display visual cards for items like weather, stocks, and sports. The company claims that in five to ten-minute test conversations, GPT-Live-1 and the mini outperform the previous Advanced Voice Mode. Significant improvements were also noted in benchmarks for scientific reasoning and agentic web searches.

      According to OpenAI, over 150 million people utilize ChatGPT’s Voice and Dictation features weekly. GPT-Live-1 will serve as the default for Go, Plus, and Pro subscribers, while the mini will be available to free users. It is compatible with iOS, Android, and the web, although the rollout might take a few days.

      The challenge of sounding human comes with its own issues. OpenAI states it has implemented voice-specific safety training and real-time safeguards that can guide responses, provide crisis resources, or terminate a conversation in higher-risk situations, with additional protections for teens. OpenAI also emphasizes that GPT-Live uses predetermined voices and will not mimic real individuals.

      This caution is crucial. A model that sounds friendly can more easily build trust and, conversely, can also mislead users. Critics, including Meredith Whittaker from Signal, have warned that chatbots “are not your friends.” The market remains competitive, with rivals like ElevenLabs and numerous voice-agent startups pursuing similar goals, prompting Amazon to revamp Alexa.

      This launch coincides with OpenAI's rollout of GPT-5.6, part of its rapid release schedule. For years, voice has been a supplementary feature rather than the primary method of interaction with ChatGPT. With GPT-Live, OpenAI is betting that speaking, rather than typing, will become the norm. Whether users desire an AI that remains constantly attentive is still an open question.