OpenAI publishes three real-time voice models to support reasoning, translation and transcription

On 8 May 2026, OpenAI officially released three real-time voice models, GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper, integrated into Realtime API. Among these, GPT-Realtime-2 has GPT-5-level reasoning capability to support interruptions and tool calls; Translate supports 70 language input to 13 outputs and synchronizes; and Whisper achieves low delayed flow. The three are charged at Token or Minutes, respectively, to address delays in voice interaction, multilingual support and real-time difficulties。