Introducing GPT-Live
GPT-Live is a new generation of voice model for natural human-AI interaction, now powering ChatGPT Voice. It offers a full-duplex architecture, enabling simultaneous listening and speaking, and intelligently delegates complex queries to advanced models like GPT-5.5 while maintaining conversation flow.
GPT-Live introduces a new generation of voice models designed for more natural human-AI interaction, now integrated into ChatGPT Voice. This innovative system features a full-duplex architecture, allowing it to listen and speak simultaneously, mimicking human conversation more closely. It can interject with acknowledgements like "mhmm" or remain silent when appropriate, making the voice experience remarkably fluid.
This model is also the most intelligent voice model to date. For intricate queries requiring web searches, deep reasoning, or complex computations, GPT-Live seamlessly delegates these tasks to advanced frontier models, such as GPT-5.5, in the background. While the processing occurs, GPT-Live continues the conversation, ensuring an uninterrupted flow. As new frontier models are released, GPT-Live will be continuously updated to leverage the latest advancements.
These enhancements significantly improve the ChatGPT Voice experience, making it more intelligent and intuitive. The research underpinning GPT-Live is expected to pave the way for increasingly complex and agentic voice-enabled tasks in the future.
Two versions, GPT-Live-1 and GPT-Live-1 mini, are currently rolling out to ChatGPT users globally, with future plans for API integration. This marks a new era in human-AI interaction, moving towards a world where collaboration with AI is as fluid and responsive as interacting with another person, with complex tasks executing seamlessly behind the scenes.
Previous voice AI systems, while foundational, had limitations. Cascaded voice systems, like the original ChatGPT Voice, relied on sequential models for speech-to-text, language processing, and text-to-speech. This approach, while groundbreaking, often led to information loss and slower, less natural responses due to the inherent complexity of chaining multiple models.
Related articles
SpaceXAI releases Grok 4.5, which Elon describes as an ‘Opus-class model’
SpaceXAI has launched Grok 4.5, an "Opus-class model" according to Elon Musk, designed to handle diverse AI tasks efficiently. This new model boasts superior token efficiency and lower costs compared to leading competitors, aiming to address rising concerns about AI operational expenses.
Data for Agents
This article explores the critical role of data in the development and improvement of AI agents. It delves into how various data types fuel agent learning, enabling them to perform complex tasks and adapt to new situations. Understanding data's impact is key to advancing AI.
Separating signal from noise in coding evaluations
OpenAI discovered significant issues in the SWE-Bench Pro coding benchmark, with an estimated 30% of tasks being broken. This highlights the challenges of creating reliable benchmarks for evaluating advanced AI models.