OpenAI announced the launch of the GPT-Live model family, intended for ChatGPT Voice, a real-time voice conversation tool. The company's goal is to make interactions sound closer to a real conversation.
OpenAI announced the launch of the GPT-Live model family, intended for ChatGPT Voice, a real-time voice conversation tool. The company's goal is to make interactions sound closer to a real conversation.
Initially, two versions will be available: GPT-Live-1 and GPT-Live-1 mini. The GPT-Live-1 model will be accessible through paid ChatGPT subscriptions, while free plan users will use the simplified version. This new feature will undergo a gradual global rollout starting this Wednesday, July 8th, and will be available on web, iOS, and Android platforms.
The GPT-Live models incorporate an architecture called 'full-duplex,' which gives them the ability to listen and speak simultaneously. According to OpenAI, this functionality allows the system to make interaction decisions multiple times per second. A notable difference is the ability not to interrupt the user if they pause to think during speech, in addition to being able to recognize subtle speech cues, such as the 'hmmm' sound used during reasoning. The artificial intelligence itself can request a moment to research data or process information before responding.
When GPT-Live identifies the need to perform complex tasks, such as web searches, reasoning, or intensive processing, it delegates these activities to GPT-5.5, OpenAI's latest multimodal model. While GPT-5.5 performs the work, GPT-Live keeps the conversation active with the user. With this, OpenAI promises significant improvements, including more natural conversations, enhanced responses, superior listening, and visual results.
ChatGPT Voice presents three levels of intelligence, where greater reasoning power requires a longer response time. In a demonstration conducted by OpenAI, GPT-Live-1 acted as an interpreter, instantly translating from English to Hindi. The promotional material also included examples of visual responses for weather forecasts, World Cup games, and maps.
Despite the advancements presented, OpenAI warned about certain restrictions. The new ChatGPT Voice may exhibit non-native accents and fluency issues in certain languages. Additionally, using the assistant in voice mode will not be compatible with activated screen or video sharing.
OpenAI announced that its newest series of powerful artificial intelligence models will be available to the general public on Thursday, pending approval from the US government.
The company's new offering, GPT-5.6, along with other advanced AI models, including Anthropic's Mythos series, has raised concerns due to their purported unprecedented ability to detect software vulnerabilities—weaknesses in code that hackers could exploit.
At the end of June, OpenAI provided preliminary access to GPT-5.6 to a limited group of partners operating only in the US, at the request of Washington.
The GPT-5.6 series includes three new models: Sol, which is the company's flagship development; Terra, a mid-level model for everyday use; and Luna, a fast and inexpensive option. The company stated in an X post on Tuesday: 'GPT-5.6 Sol, along with Terra and Luna, will be publicly released this Thursday. We are expanding early access globally now,' without providing further details.
According to the American publication Axios on Tuesday, citing a source familiar with the situation, the Trump administration permitted OpenAI's wide release of GPT-5.6 after testing and meetings between the company and government officials.
AFP contacted OpenAI, the White House, and the US Department of Commerce for comments regarding the Axios report. This move comes against a backdrop of a similar situation with OpenAI's main competitor, Anthropic, which last week announced the resumption of global access to its most powerful AI models, Fable 5 and Mythos 5, after the US government lifted restrictions on their release.
