OpenAI dropped three new voice AI models on May 7: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. Together they handle real-time conversation, live translation across more than 70 languages, and low-latency transcription.
What the new models actually do
GPT-Realtime-2 is the flagship of the bunch, sitting within OpenAI’s GPT-5 tier. It packs a 128K context window, which means it can hold the equivalent of a short novel’s worth of conversation in memory during a single session.
The model scored a 15.2% improvement on the Big Bench Audio benchmark compared to its predecessors.
GPT-Realtime-2 supports parallel tool calls, adjustable reasoning effort, and what OpenAI describes as “strong recovery behavior.”
