GPT Realtime Mini

GPT Realtime Mini is OpenAI's smaller, cheaper speech-to-speech model for the Realtime API, designed for high-volume and latency-sensitive voice applications. It does native speech-to-speech in one model flow, cutting the delay of pipelined STT-plus-LLM-plus-TTS systems, and reacts to pauses, interruptions, and noise changes mid-conversation. A December 2025 snapshot improved word-error rates on real-world audio, reduced hallucinations during silence, and made voice output more natural.

It trades some reasoning capability for lower cost and faster responses relative to the flagship GPT Realtime line, and is the recommended default for production voice agents where per-minute cost matters. OpenAI describes the 2.1-mini variant as a distilled reasoning model priced far below the full model, with support for WebRTC, WebSocket, and SIP connections.