Rate Limits and Quotas: Why AI Tools Sometimes Say Slow Down

Every AI API and subscription caps how much you can use in a given time. Hit the ceiling, and you get a 429 instead of an answer.

Why limits exist at all

Running a large model costs real computing power. If one user could send unlimited requests per second, they could overload the service for everyone else, or run up a bill nobody expected. So providers cap usage - a rate limit caps how much you can send per minute, a quota caps how much you can use over a longer period like a day or a month.

What a rate limit actually looks like

Claude's API, for example, limits requests per minute, input tokens per minute, and output tokens per minute - separately, per model. Cross any one of them and the next request fails, even if the other two limits still have room left.

The error you'll see: 429

When you go over a limit, the API answers with HTTP status code 429, known as 'Too Many Requests.' The response often includes a retry-after header telling you exactly how many seconds to wait before trying again - so the fix is usually built right into the error.

Coping with limits in practice

Wait and retry: back off for a moment, ideally with a growing delay, instead of hammering the API again immediately. Batch: combine several small requests into fewer, larger ones where the API supports it. Downsize: route simple, high-volume tasks to a smaller, cheaper model with its own separate limit, saving your budget on the strong model for what actually needs it.

EXAMPLE

A 429 response you might see: status 429, header retry-after: 12 - meaning wait 12 seconds before sending your next request instead of retrying immediately.

🛠️ EXERCISE — TRY IT YOURSELF

Find and read the rate limit headers on a real API response.

Make an API call to Claude or OpenAI, or check existing code/logs that call one.
Look at the response headers for rate-limit fields, e.g. anthropic-ratelimit-requests-remaining or the OpenAI equivalent.
Note how many requests or tokens you have left before hitting the limit.
Optional: carefully trigger a 429 by sending a handful of fast requests and read the retry-after value - note the successful requests before it still cost tokens, so prefer a test key and stop after a few tries.

✅ SELF-CHECK

☐ Does your code wait longer after a 429 error instead of retrying immediately?
☐ Does your code have a fixed upper limit on the number of retries?
☐ Does your code use the retry-after header instead of assuming a fixed wait time?

QUICK QUIZ

You keep getting HTTP 429 errors from an AI API. What's the recommended way to handle it?