OpenAI Unveils GPT-Live, a New Voice Model That Can Listen and Talk at the Same Time

OpenAI announced Wednesday the launch of GPT-Live, a new generation of voice models for ChatGPT that enables natural, continuous real-time conversation. The technology, which begins rolling out today to all ChatGPT users worldwide, replaces the previous voice mode and represents a fundamental shift in how voice interaction with artificial intelligence functions.

The launch was unveiled by company CEO Sam Altman and co-founder Alex Kraus at a public event. OpenAI is releasing two versions of the model: GPT-Live-1, which will serve as the default for paid subscribers, and GPT-Live-1 mini, designated for free-tier users. Both models supersede their predecessors and power the redesigned ChatGPT Voice experience.

At the core of the technology is a Full-Duplex architecture, which enables the system to listen and speak simultaneously. Unlike earlier voice models — which had to wait for the user to finish speaking before generating a response — GPT-Live processes speech in real time and continuously determines whether to continue responding, pause, listen, or invoke an external tool, depending on how the conversation evolves.

For users, the practical result is a markedly smoother and more natural conversational experience. The model can be interrupted mid-sentence without losing conversational flow, handles background noise and natural speech hesitations, and can emit brief acknowledgment cues to signal active listening. It also supports real-time translation while the user is still speaking, and can comply with a request to remain silent until directly addressed — a capability absent from earlier versions.

A notable architectural feature is the separation between managing the voice conversation and executing more computationally intensive tasks. When a user poses a question requiring web search or advanced processing, GPT-Live offloads the task in the background to the company's flagship model — GPT-5.5 at launch. While the background model performs the computation, GPT-Live continues managing the live conversation uninterrupted. OpenAI noted that the background model will be updated continuously as more capable models are released.

The system also introduces three selectable reasoning tiers: Instant for rapid responses, and Medium and High for tasks demanding deeper analytical processing, drawing on the company's reasoning model suite.

In a further enhancement, the upgraded ChatGPT Voice will be capable of rendering visual cards on screen during a conversation. For topics where graphical presentation adds clarity — weather forecasts, financial market data, sports results — visual components will appear alongside the spoken response.

OpenAI also published internal evaluation data. GPT-Live-1 was preferred by 75.7% of evaluators over the previous voice mode, while GPT-Live-1 mini achieved a preference rate of 69.2%. On the GPQA scientific reasoning benchmark, GPT-Live-1 reached an accuracy rate of 84.2%, compared with 45.3% recorded for the prior voice mode — a substantial performance gap.

The model incorporates real-time safety mechanisms operating in parallel with the conversation. When the system detects potentially harmful output, it can redirect the model toward a safer response, surface relevant support resources, or terminate the conversation outright in high-risk scenarios — including content related to self-harm. For younger users, the model applies age-adapted behavioral parameters, with parental oversight remaining available through existing controls. Built-in safeguards also prevent the model from replicating the voices of real individuals.

The broader rollout of GPT-Live is expected to expand progressively through 2026. GPT-Live-1 will serve as the default voice model for Plus, Pro, and Go subscribers, while GPT-Live-1 mini will be available to users on the free tier.