What are considered "requests" when using the real-time API?

API
,
1

I’m looking at the rate limits for the real-time API models, specifically the new transcription models:

https://platform.openai.com/docs/models/gpt-4o-mini-transcribe
https://platform.openai.com/docs/models/gpt-4o-transcribe

They list the Tier 5 limit as 10,000 RPM.

What is considered a request for the real-time API? Is it the number of WebSocket connections? Or is it the number of requests sent over those WebSocket connections?

Assuming we went audio in 300ms chunks to the endpoint, that’d be 200 RPM per user, so the limit would be 50 concurrent users.

Or is the limit 10,000 concurrent users?

2

Hi @thomasb
Did you find answer?

3

Despite the model being “realtime”, that is just a buffering front end.

The model generates based on a set input context and a trigger. The trigger is either the end of server voice activity detection, or your response.create event. That is your request to receive a generated output, and what the rate limiter would error out on if over the limiter quota.