GPT-Realtime-2 is a native speech-to-speech AI model developed by OpenAI, designed for low-latency, high-reasoning voice interactions. Unlike traditional voice pipelines that stitch together separate transcription, language, and synthesis models, GPT-Realtime-2 operates as a single end-to-end architecture. This allows it to process raw audio input and generate spoken responses directly, preserving non-verbal cues such as tone, pacing, and emotional nuance that are often lost in text-based conversions.
The model introduces "GPT-5-class" reasoning to real-time audio, enabling it to handle complex, multi-turn conversations and nuanced instructions. A key feature is the adjustable reasoning effort parameter, which allows developers to control the intensity of the model's cognitive processing across levels such as minimal, low, medium, high, and xhigh. This flexibility enables the optimization of responses for either immediate speed or deeper logical accuracy depending on the complexity of the user's request.
New Audio Features
To improve the user experience in live settings, GPT-Realtime-2 includes a Preambles system. This allows the model to emit a "commentary" phase—short, contextually relevant verbal fillers—while it continues to process more intensive reasoning in the background. This technique effectively reduces perceived latency and bridges the silence during complex computations. Additionally, the model features enhanced interruption handling and more reliable tool use, allowing voice agents to perform external tasks (like checking a CRM or booking a flight) while maintaining the flow of a live conversation.
Architecture and Capabilities
GPT-Realtime-2 supports a 128,000-token context window (with some enterprise variants supporting up to 256,000 tokens) and can generate up to 32,000 output tokens. It is optimized for high-reliability workflows where strict instruction following is required. The model was released as part of a specialized audio suite that includes GPT-Realtime-Translate for live speech-to-speech translation and GPT-Realtime-Whisper for streaming transcription.
For optimal performance, official guidance suggests defining clear confirmation boundaries and specific decision points within the system prompt. Rather than using generic instructions like "be concise," developers are encouraged to define what conciseness means in the context of the specific task—such as providing direct answers for tool results while offering more detail for troubleshooting scenarios.