OpenAI Releases Three Real-Time Voice Models to Enhance Smooth Voice Interaction and Cross-Language Communication

OpenAI recently released three real-time voice models integrated into the Realtime API for developers, aiming to solve pain points like latency, interruption handling, and multi-language support in voice interactions. These models target reasoning, translation, and transcription scenarios, providing underlying technical support for real-time voice applications.

Among them, GPT-Realtime-2 is the first voice model with GPT-5-level reasoning capabilities. It can handle user interruptions or corrections while maintaining smooth conversations and supports tool calls, helping developers build complex voice assistants. GPT-Realtime-Translate supports 70 input languages to 13 output languages, with translation speed synchronized to the speaker. GPT-Realtime-Whisper focuses on low-latency streaming transcription, suitable for real-time subtitles and other scenarios.

In terms of pricing, GPT-Realtime-2 is charged by Token: $32 per million tokens for input, $64 for output. The translation and transcription models are charged per minute: $0.034 and $0.017 respectively.