gpt-realtime-2

The GPT Realtime 2.x series is a speech-to-speech model line with built-in reasoning, accepting audio input and producing audio output rather than transcribing to text along the way. Azure-hosted documentation frames the family as designed for low-latency, interactive voice experiences where stronger instruction following and on-the-fly reasoning matter more than in earlier realtime models, making the architecture a fit for conversational agents, voice-driven assistants, and other scenarios that need responsive spoken exchange with the model thinking through user intent before responding.

Beyond the core audio-in, audio-out design, the 2.x preview introduces a configurable reasoning effort control that lets developers tune how much deliberation the model invests per turn, paired with response phases that separate internal commentary from the final spoken answer so downstream systems can distinguish planning from delivered output. A substantially expanded 256,000-token context window also positions the family for longer, multi-turn dialogues where earlier realtime models would have run out of room, supporting richer conversational memory and more sustained interactive sessions.