OpenAI has introduced GPT Realtime, a model designed to power real-time, spoken conversations inside ChatGPT. The production-ready successor to ChatGPT’s Standard Voice and Advanced Voice modes is now rolling out to users, replacing the older text-in-the-middle pipeline with end-to-end audio handling for more natural, lower-latency voice interactions.<\/p>\n
GPT Realtime is built specifically for voice. The model accepts audio directly as input, processes it, and returns audio as output without first transcribing speech to text, an approach OpenAI describes as more natural and more responsive than the pipeline used by earlier voice features.<\/p>\n
Key characteristics highlighted in the launch include:<\/p>\n
For everyday users, the practical change is a more fluid voice mode in ChatGPT. Users can interrupt the assistant, change topics quickly, and ask the model to respond in a different language without restarting the session. The voice that has been associated with Advanced Voice remains available as a default option.<\/p>\n
For developers, GPT Realtime is accessible through the Realtime API in the OpenAI platform. The API supports the same audio-to-audio interface that powers the in-app experience, including streaming, function calling, and MCP integration, so applications built on the API can match the capabilities users see inside ChatGPT.<\/p>\n
Voice has been one of the more challenging modalities for large language models because conversational speech requires both low latency and the ability to handle the ambiguity, emotion, and pace of natural talk. Earlier voice modes addressed this by routing audio through a transcription model, a text-based language model, and a speech synthesis model, which added delay and could lose acoustic cues. End-to-end audio models are an attempt to solve that by letting a single model reason over the raw sound.<\/p>\n
OpenAI’s announcement positions GPT Realtime as a step in that direction, with measurable gains in responsiveness and voice quality over the previous Standard and Advanced Voice options. Pricing for the Realtime API is published separately on OpenAI’s developer pricing page, and the company notes that usage-based audio input costs have been reduced.<\/p>\n
The launch suggests several follow-on questions for the year ahead: how third-party developers adopt the Realtime API, how voice agents perform on longer and more complex tasks, and how competitors respond with their own end-to-end speech models. For now, GPT Realtime sets a new baseline for what users should expect from a voice assistant in a chat product.<\/p>\n
GPT Realtime is OpenAI’s production-ready model for real-time, spoken conversations inside ChatGPT. It is the successor to ChatGPT’s Standard Voice and Advanced Voice modes and processes audio end-to-end without first transcribing speech to text.<\/p>\n
Earlier voice modes routed audio through a speech-to-text model, a text-based language model, and a speech synthesis model, which added latency and could lose acoustic cues. GPT Realtime handles audio directly as input and output, and adds improved function calling, MCP support, and mid-conversation language switching.<\/p>\n
Developers can access GPT Realtime through the Realtime API in the OpenAI platform. The API supports streaming audio, function calling, and MCP integration, and pricing is published on OpenAI’s developer pricing page, where usage-based audio input costs have been reduced.<\/p>\n