OpenAI puts GPT-Live into ChatGPT Voice
The new full-duplex voice models can listen while speaking, delegate harder work to GPT-5.5 and will come to the API later.
By Ryan Merket · Published
Why it matters
GPT-Live moves ChatGPT Voice from turn-based chat toward a live agent interface, with API access pending and GPT-5.5 handling harder work in the background.
OpenAI launched GPT-Live on July 8th, replacing the core of ChatGPT Voice with a pair of full-duplex voice models designed to listen and speak at the same time.
The rollout starts with two models, GPT-Live-1 and GPT-Live-1 mini. OpenAI says GPT-Live-1 will become the default voice model for ChatGPT Go, Plus and Pro users, while GPT-Live-1 mini will become the default for Free users. The models are rolling out globally across iOS, Android and ChatGPT.com, and OpenAI says API access will follow. Developers and enterprises can register through OpenAI's notification form.
The shift is architectural. GPT-Live is built to process audio input continuously while generating audio output, which means the model can decide many times per second whether to speak, pause, keep listening, interrupt or call a tool. OpenAI says that lets ChatGPT Voice acknowledge a speaker with short cues such as "mhmm" or "yeah," respond in faster back-and-forth exchanges and stay quiet when the user is thinking.
GPT-Live also separates the conversational layer from heavier reasoning work. When a request needs search, deeper reasoning or agentic work, GPT-Live can delegate the task to a frontier model in the background and return the answer through the live conversation. At launch, OpenAI says GPT-Live uses GPT-5.5 behind the scenes. GPT-Live-1 Instant and GPT-Live-1 mini use GPT-5.5 Instant, while GPT-Live-1 Medium and GPT-Live-1 High use GPT-5.5 Thinking with medium and high reasoning effort, according to the company.
That decoupling is the real product decision. OpenAI is trying to make voice feel immediate without forcing every spoken exchange through the slowest available reasoning path. The live model handles timing, interruptions and conversational flow. The frontier model handles the work that benefits from more compute. If it performs as advertised, ChatGPT Voice becomes less like dictation into a chatbot and closer to an interface that can manage long-running tasks while the user keeps talking.
OpenAI framed GPT-Live against two earlier voice systems. The original ChatGPT Voice, announced on September 25th, 2023, used a cascade of three models: speech-to-text, a language model and text-to-speech. OpenAI said that design made early voice access possible, but introduced latency and information loss between models. GPT-4o, announced on May 13th, 2024, moved audio, text and vision into a single model and lowered latency, but OpenAI's new post says Advanced Voice Mode still worked through discrete turns and depended on silence-based turn detection.
GPT-Live is OpenAI's attempt to fix that last mile of voice. Silence detection is a small technical detail with large product consequences: a pause can mean the user is done, thinking or being interrupted by background noise. OpenAI says GPT-Live can better distinguish those cases because it continuously processes input and output rather than waiting for a turn to end.
The company is leaning on internal human evaluations to make the case. In matched 5-to-10-minute conversations against Advanced Voice Mode, OpenAI says GPT-Live-1 was preferred 75.7% of the time and GPT-Live-1 mini was preferred 69.2% of the time. OpenAI also reported mean conversation-flow ratings of 4.96 for GPT-Live-1, 4.33 for GPT-Live-1 mini and 3.80 for Advanced Voice Mode, on a 7-point scale. For pleasantness, it reported 5.19 for GPT-Live-1, 4.47 for GPT-Live-1 mini and 3.82 for Advanced Voice Mode.
Those figures are company-run evaluations, so they should be read as OpenAI's product evidence rather than independent benchmarking. The more consequential number in the announcement is usage: OpenAI says more than 150 million people each week talk to ChatGPT through Voice and Dictation. Voice is already a mass interface for the company, which makes the default-model change immediately material for ordinary ChatGPT use.
The new ChatGPT Voice experience also adds visual responses while a conversation is underway. OpenAI says voice sessions can show cards for weather, stocks, sports and other topics, and that Voice continues to support search, memory, images and file uploads. The result is a voice interface that can speak, listen and display structured information in the same session.
There are limits at launch. GPT-Live will support ChatGPT Voice across mobile and web, but OpenAI says it will not initially support voice with video or screen sharing in ChatGPT. Users can still access legacy versions of ChatGPT Voice, including Standard and Advanced Voice Mode, where those features are available.
OpenAI is also positioning the release around voice-specific safety, a necessary move for a system designed to hold longer, more natural conversations. The company says it expanded testing with audio-native evaluations and synthetic audio evaluations covering self-harm, psychosis and mania, emotional reliance on AI, violence and sexual content. OpenAI also says GPT-Live includes safeguards that can act while the model is speaking, including steering unsafe output, surfacing safety messages or ending higher-risk conversations.
For teen users, OpenAI says parents can choose whether ChatGPT Voice is available through Parental Controls, and linked parents may be notified in higher-risk situations involving signs of potential self-harm or suicidal intent. The company also says GPT-Live uses predefined ChatGPT voices and includes safeguards against imitating a real person's voice.
The API is the next test. In ChatGPT, OpenAI controls the user experience, the voices, the safety layer and the default model routing. In the API, developers will want to embed live voice into customer support, tutoring, coaching, sales and workplace agents. That puts pressure on pricing, latency, reliability and policy enforcement in settings OpenAI does not fully design.
GPT-Live gives OpenAI a cleaner way to sell voice as an application platform. The model can keep a conversation alive while deeper work happens elsewhere in the stack. The company now has to prove that the architecture holds up outside polished demos and controlled evaluations, especially once developers start building live agents that talk over messy audio, handle pauses incorrectly and work in environments where a bad interruption can break the task.