OpenAI Rolls Out Three Voice Models, Bringing GPT-5 Reasoning to Conversation

On May 7, 2026, OpenAI unveiled three new voice models for its developer-facing Realtime API. Anchored by GPT‑Realtime‑2, which builds GPT-5-class reasoning directly into speech processing, the lineup adds dedicated models for live translation across more than 70 languages and for transcribing while a speaker is still talking, sharpening the company's strategy of making voice AI's primary interface.