OpenAI releases GPT-Realtime-2 reasoning voice model with two specialized variants for translation and transcription

OpenAI has released three new realtime voice models through its Realtime API, with the flagship GPT-Realtime-2 incorporating what the company describes as "GPT-5-class reasoning" capabilities.

GPT-Realtime-2: Voice model with reasoning

GPT-Realtime-2 is designed for live voice interactions where the model maintains conversation flow while processing complex requests. According to OpenAI, the model can "reason through a request, call tools, handle corrections or interruptions, and respond in a way that fits the moment." The model is priced at $32 per 1M audio input tokens ($0.40 for cached input tokens) and $64 per 1M audio output tokens.

GPT-Realtime-Translate: Live speech translation

GPT-Realtime-Translate provides real-time speech translation from 70+ input languages into 13 output languages. The model is designed to maintain pace with the speaker during live translation. Pricing is set at $0.034 per minute.

GPT-Realtime-Whisper: Streaming transcription

GPT-Realtime-Whisper offers low-latency streaming transcription that processes speech as users speak. OpenAI positions this model for applications requiring live captions and real-time meeting notes. The model costs $0.017 per minute.

Availability and implementation

All three models are now available through OpenAI's Realtime API. Developers can test the models in OpenAI's Playground. The company has not disclosed specific benchmark scores, context window sizes, or training data cutoff dates for these models.

What this means

The release of GPT-Realtime-2 with claimed "GPT-5-class reasoning" represents OpenAI's first voice model with advanced reasoning capabilities, potentially enabling more sophisticated voice-based applications beyond simple command-and-response patterns. The specialized translation and transcription models address specific use cases with per-minute pricing that may be more predictable for developers building streaming applications. However, without published benchmarks or technical specifications, the actual performance improvements over existing voice models remain unclear.