gpt-realtime - For reliable, production-ready voice agents

Hi everyone!

OpenAI's new gpt-realtime model is big step forward for voice agents. The key isn't just a faster model, but a shift in how it understands.

For a true voice agent to work, it needs to understand the subtle cues in our speech, the tone, the pauses, the emotion. That's what carries the real meaning. gpt-realtime is built on a voice-in, voice-out approach. It processes audio directly, without first transcribing it to text. This is the direction the field has been trying to break through.

Also great to see the Realtime API is now generally available, with practical new features for production like remote MCP server support and SIP integration.

So cool! Now companion products can integrate with the Realtime API, which is a big step forward for improving user experience. I can't wait to try out real-time conversations! @OpenAI

Voice is definitely faster than typing. Is this the end of open-landscape offices?

We just have to wait a littlebit more and we can communicate with the ChatGPT right while driving, without looking at the iphone screen...

This looks amazing — love how you’re empowering creators to scale AI experiences.

The real test will be, can it pick up hesitation, sarcasm, or subtle emphasis? That’s where most AI agents break down.

Pretty cool update