OpenAI Cut Realtime API Prices 60% the Day Before My Demo

Built a realtime voice client on the OpenAI Realtime API. Spent the time, got it working, told someone there'd be a demo tomorrow.

Today OpenAI dropped prices 60% and added WebRTC support.

The demo is still happening. The client still works. But the version of the story where I say "and this is why it was worth the engineering effort" now has a different shape — because the engineering effort just got cut in half retroactively, by a company that had no idea I was going to demo this and simply does not care.

WebRTC is the interesting part. The Realtime API launched on WebSockets, which is fine, WebSockets are fine, but WebSockets require you to care about a lot of things — reconnection, state, the specific flavor of misery that comes from a dropped connection mid-sentence when the model is halfway through a thought. WebRTC handles more of that at the transport layer, and more importantly it's the thing that every browser already knows how to do for voice. The path from "user clicks button" to "audio is flowing" just got shorter.

The 60% price cut is almost beside the point — almost. Real-time voice inference is expensive in a way that matters for what you can actually build. You can run a demo at any price. You cannot run a product at any price. That number moving means the product conversation opens up, and the product conversation is the only one worth having.

So: good timing. Embarrassing timing. The kind of timing that makes you wonder if you should have waited a day to build it, and then you remember you had no way of knowing, and neither did anyone else, and that's just the job.

Demo's tomorrow. The code costs less than it did this morning.