Outage or rate limit? How to tell the difference

They look alike from the outside: the answer does not arrive. But one is a problem on the provider’s side and the other is a limit on your account. The fix is different.

Start with the message

A message about limits, usage, credit or billing points to your account. Providers document limits as rate limits, spend caps and used-up credit. Nothing is broken.
A message about being overloaded, unavailable or timed out points to the service. Providers describe these as temporary and suggest retrying after a wait.
A bare error number. Use the guide to error codes to see which side it belongs to. As a rule, 429 is about limits and 500, 503 and 529 are about the service.

Then check whether it is everyone

Open the provider’s status page and ours. If either reports a problem with the part you are using, you can stop and wait.
If both look healthy, try another route in: the other app or the website, a different browser, a different network.
If you use an API, try a different key, project or model if you have one. A fault that follows your account is probably a limit. A fault that follows the model is probably the service.

Why this is not always clear-cut

Some limits are sensitive to how fast you ramp up. Anthropic notes that a sharp rise in an organisation’s usage can bring 429 errors because of what it calls acceleration limits, and advises increasing traffic gradually. OpenAI documents a “slow down” response for the same reason. So a 429 can follow a sudden jump in your own traffic, and a mix of 429s and 5xx errors does not by itself prove which problem you are seeing.

What not to do

Do not hammer it. Perplexity’s documentation warns that aggressive retry loops without delays make throttling worse. Waiting longer between tries is the standard advice.
Do not keep retrying errors about your request or your account. A wrong key or an empty balance will fail every time.

When it really is the provider

If the provider reports the problem, there is little to do but wait and have a fallback. Our checklist for developers covers how to build that in. For everyday users, what to do when an AI chatbot stops working has the practical steps.

Sources

Read on 8 October 2026. Providers change their documentation, so check it for anything important.