OpenAI Ships Voice Models That Reason As You Speak

OpenAI announced a new family of realtime voice models on May 7, the first time the company has put its full reasoning stack behind audio in the API.

The headline release is GPT-Realtime-2, which OpenAI describes as its first voice model with GPT-5-class reasoning. It can handle harder requests, call tools mid-conversation, recover from interruptions, and stay coherent when a user corrects it. Two smaller companion models arrived alongside it: GPT-Realtime-Translate, which carries live speech across 70 input languages and 13 output languages, and GPT-Realtime-Whisper, a streaming transcriber for live captions.

The pitch to developers is that voice apps no longer have to choose between speed and intelligence. Earlier realtime audio models held conversations at low latency by leaning on a smaller, less capable backbone. Reasoning either happened in a separate text model or not at all. GPT-Realtime-2 keeps latency low while letting the model think before it answers, and it tracks context across a 128,000 token window, four times the previous version.

On Big Bench Audio, OpenAI reports a 15.2 percent improvement over GPT-Realtime-1.5 at the high tier. Pricing reflects the new positioning. GPT-Realtime-2 costs $32 per million audio input tokens and $64 per million audio output tokens, with cached input falling to $0.40.

Voice has been a quieter front in the model wars than reasoning or coding, but the practical stakes are clear. Live translation that keeps pace with the speaker, and streaming captions that arrive as the speaker talks, change what an accessibility product or contact centre can build at API prices. The competitive context is just as crowded: Anthropic recently put ten finance agents inside Claude, and the broader race over frontier capability is now drawing government attention.

The new models are live in the OpenAI API. Early hands-on coverage is available from 9to5Mac and The Next Web.