I'm trying to set a hard spending limit in dollars for Azure OpenAI/Foundry. I don't want to place an API gateway in front of it; I'd prefer the service or resource to stop processing requests once the limit is reached. My current idea is to use a Lambda or similar function to monitor usage logs, calculate input/output token costs, and then disable network access or the OpenAI resource when the budget is exceeded. Is there a better or more reliable approach?
4 Answers
Be careful not to confuse a billing alert with an actual spending cap. An alert only notifies you; it does not necessarily stop consumption. Credits and account-specific billing rules can also affect how caps work, and charges may continue during a grace period even after a limit is reached. Verify the exact terms for your subscription before relying on it as a hard cutoff.
A subscription spending limit may sound relevant, but it generally won’t give you a clean way to disable only the Azure OpenAI or Foundry resources. You’d likely need to automate a more targeted action, such as disabling the OpenAI resource or its deployments when the threshold is reached.
There is currently no obvious built-in Foundry spending control at the project or deployment level. A custom monitor could calculate estimated input and output costs, then disable the resource or block its network access. Just account for logging and billing delays, because the cutoff probably won’t be perfectly instantaneous.
I agree that a custom token-and-cost calculation with an automated resource block is probably the most realistic short-term solution. I’d rather stop the whole resource than risk allowing usage to continue after the threshold.
There doesn’t appear to be a native Foundry control that automatically stops usage at a dollar amount. The practical option is to combine Azure Cost Management budget alerts with automation. For example, trigger an Azure Function or webhook when actual or forecasted spending reaches a threshold, then disable the relevant resource, deployments, or access path.
That makes sense. Disabling the entire OpenAI resource once usage reaches the threshold may be the safest approach, with the limit calculated from token usage logs since they seem to provide the closest thing to near-real-time tracking.