OpenAI Releases gpt-realtime-2.1 and gpt-realtime-2.1-mini — 25% Lower Latency and Reasoning for Voice Agents

What's new

• gpt-realtime-2.1: Improved alphanumeric recognition, noise and interruption handling; 25% lower p95 latency via improved caching
• gpt-realtime-2.1-mini: Smaller reasoning model for real-time voice applications at 6× lower cost than the full version
• Smart Turn Detection: ML model predicts end-of-turn at silence boundaries for more natural dialogue
• Improved robustness: Better resilience to noise, interruptions, and stream restarts

Why it matters

Realtime models are critical for customer support, AI phone assistants, and voice-first applications. The mini variant with reasoning capabilities opens the door to complex logic in voice agents at low cost.

How to try it

Available via the OpenAI Realtime API as gpt-realtime-2.1 and gpt-realtime-2.1-mini; test in the Realtime section of the OpenAI Playground.