I’m seeing different request and token limits in the OpenAI dashboard and in the rate-limit headers returned by the API. I’m not sure whether I should treat the headers as the live limit for a particular request, or whether the dashboard’s project and organization settings should take precedence.

For a client that serves several users, which values should drive throttling, and how do you avoid backing off too aggressively when only one limit is close? I’m especially interested in how others interpret rate-limit headers when limits change or multiple projects are involved.