OpenAI adds 3 new audio models to the Realtime API

OpenAI has added GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper to the Realtime API.

GPT-Realtime-2 is a voice model with GPT-5-level reasoning, supporting tool calling during conversation, handling interruptions and corrections, short preambles, and more natural tone control. The context window has grown from 32K to 128K, and reasoning effort can be chosen among minimal / low / medium / high / xhigh, with low as the default.

Improvements include the following.

• parallel tool calls and tool transparency
• Better recovery behavior
• Stronger domain understanding
• More refined tone and delivery

OpenAI presented three patterns—voice-to-action, systems-to-voice, and voice-to-voice—and emphasized real-world applicability by citing cases from Zillow, Deutsche Telekom, Priceline, Vimeo, and BolnaAI. In terms of performance, GPT-Realtime-2 (high) scored 15.2% higher than GPT-Realtime-1.5 on Big Bench Audio, and xhigh scored 13.8% higher on Audio MultiChallenge.

The models released alongside it are as follows.

• GPT-Realtime-Translate: real-time translation from 70+ input languages into 13 output languages
• GPT-Realtime-Whisper: streaming STT that generates captions and transcripts as soon as speech occurs

Pricing for GPT-Realtime-2 is $32/1M tokens for audio input, $0.40/1M tokens for cached input, and $64/1M tokens for audio output, while Translate is $0.034/minute and Whisper is $0.017/minute. All three models are available in the Realtime API and Playground, and OpenAI also presented safety measures including active classifiers and Agents SDK guardrails.