OpenAI opens three real-time voice models

OpenAI’s realtime voice push is not a June 1 announcement. The company published the release on May 7, 2026, in a product post describing three new audio models available through its API for developers building live voice applications. OpenAI said the models are designed to let apps “reason, translate, and transcribe speech” as people speak, extending the company’s existing Realtime API rather than creating a separate consumer product. (openai.com) The three models named by OpenAI are GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Transcribe. In OpenAI’s description, the first is positioned as a speech-to-speech model for harder conversational tasks, the second as a live translation model, and the third as a transcription model for turning spoken audio into text. (developers.openai.com) ### Which three models did OpenAI actually release? OpenAI’s May 7 post says GPT-Realtime-2 is its “first voice model with GPT-5-class reasoning,” aimed at requests that need better context handling and more natural back-and-forth. The company’s developer guide separately describes gpt-realtime-2 as its state-of-the-art reasoning voice model for low-latency speech-to-speech applications. (openai.com) GPT-Realtime-Translate is described by OpenAI as a live translation model. The company said it can translate speech from more than 70 input languages into 13 output languages while keeping pace with the speaker. (openai.com) GPT-Realtime-Transcribe is the transcription-focused model in the set. OpenAI’s realtime documentation says transcription sessions are meant for applications that need text from audio without model-generated spoken responses. ### Is this a brand-new voice API or an expansion of existing tools? OpenAI already had a Realtime API before this release. (openai.com) The company introduced that API in public beta in 2024 for low-latency multimodal experiences and later expanded it with production features and the gpt-realtime model. The May 2026 release adds a new generation of realtime voice models on top of that existing framework. (openai.com) OpenAI’s current developer docs show the same structure. Voice-agent sessions connect to `/v1/realtime`, accept audio or text, and return model responses, tool calls and session events, while separate translation and transcription sessions serve narrower use cases. (developers.openai.com) ### What can developers do with them in practice? OpenAI’s documentation says developers can use the models for low-latency speech-to-speech applications, live interpretation and streaming transcription. The company also says realtime models support tool use, which lets a voice agent call external functions while handling a conversation. (openai.com) The company’s broader audio guide says audio applications can mix spoken input, spoken output and text generation in one interaction. That means the release is aimed at builders of customer-service agents, language tools, assistants and other applications where users speak and expect an immediate spoken or textual response. (developers.openai.com) ### Why did some reports make it sound like this happened on June 1? OpenPR published an item on June 1, 2026, that referred to OpenAI opening realtime voice to developers. But OpenAI’s own product page dates the release to May 7, 2026, and its release index lists the same item under that date. The June 1 story appears to have been describing an already published OpenAI launch. (developers.openai.com) ### Where do developers find the next step? OpenAI’s next step for developers is in its platform documentation, not a future event date. The company points users to its realtime model prompting guide, realtime overview, and audio learning pages for implementation details and examples. (developers.openai.com 1) (developers.openai.com 2) (openai.com)