API rate limits explained: tiers, 429 errors and how to avoid them

API rate limits cap how many requests and tokens your account can send to a model per minute or per day, and when you exceed one the API answers with an HTTP 429 error instead of a response. On October 23, 2026 OpenAI, Anthropic and Google all use automatic usage tiers that raise your limits as you spend more, but they count differently: Anthropic ignores cached input and max_tokens in its limits, OpenAI counts max_tokens, and Google adds a spend limit per 10 minutes. This guide explains how each one works, what a 429 means and how to stop hitting it, based on each vendor's documentation read on October 23, 2026.

What gets counted

Every vendor measures several things at once, and you hit the limit on whichever runs out first:

• RPM: requests per minute.
• TPM: tokens per minute. Anthropic splits it into input (ITPM) and output (OTPM).
• RPD and TPD: requests and tokens per day, on some models and tiers.
• IPM: images per minute, for image models.

Limits apply to your organization or project, not to each API key or user. Google's example makes the point: with a limit of 20 RPM, the 21st request in a minute fails even if you have plenty of tokens left.

Tiers by vendor

Table: usage tiers (official documentation, checked October 23, 2026)

For a sense of scale, OpenAI's Build tier allows 5,000 requests and 1,000,000 tokens per minute on its Astra, Sol and Terra models, and Anthropic's Start tier allows 1,000 requests, 2,000,000 input tokens and 400,000 output tokens per minute on Claude Opus 5.5. Our guides to OpenAI API pricing and Claude API pricing cover what those tokens cost.

The details that change your design

• Anthropic does not count cached reads. For most Claude models, only uncached input and tokens written to the cache count toward ITPM. Anthropic's example: with a 2,000,000 ITPM limit and an 80% cache hit rate, you can process 10,000,000 input tokens per minute. Prompt caching is therefore a throughput tool as well as a discount.
• max_tokens counts at OpenAI, not at Anthropic. OpenAI calculates your usage as the larger of max_tokens and the estimated tokens of the request, so set it close to the size you expect. Anthropic counts only the output tokens actually generated, so a high max_tokens has no rate limit cost there.
• Limits are enforced over short windows. Anthropic uses a token bucket that refills continuously and warns that 60 RPM may be enforced as 1 request per second, so a burst can fail even under the per-minute number.
• Sudden growth is limited too. OpenAI's slow_down error and Anthropic's acceleration limits trigger when traffic rises too fast. OpenAI's rule of thumb: once you reach 1 million TPM, grow by no more than 50% every 15 minutes.

Not every 429 is a rate limit

A 429 can mean you sent too much too fast, and waiting fixes it. But it can also mean you reached a spend limit, and then no amount of retrying helps. When an Anthropic organization reaches its tier's monthly spend cap, the API returns 429 with no retry-after header until 00:00 UTC on the first day of the next month, unless you request a higher limit. OpenAI returns 429 for a hard spend limit you configured, and its documentation says Retry-After does not mean that quota or billing errors can be solved by retrying. Google returns 429 RESOURCE_EXHAUSTED for its per-10-minute spend limit. Read the error body before retrying.

How to stop hitting the limit

1. Respect Retry-After. OpenAI and Anthropic send it on temporary 429s. If it is missing, use exponential backoff with random jitter and cap both the attempts and the total time. OpenAI warns that failed requests also count toward your per-minute limit, so resending immediately makes it worse.
2. Let the SDK retry, but only once. OpenAI's official SDKs retry eligible 429 and 503 errors automatically; if you add your own retries, disable the SDK's or count them.
3. Read the headers. OpenAI returns x-ratelimit-remaining-requests and x-ratelimit-remaining-tokens; slow down before you reach zero.
4. Move bulk work to batches. The batch API runs large jobs without using your real-time limits, at a discount.
5. Cache what repeats, especially on Claude, where cached input does not count.
6. Ramp up gradually after a launch instead of switching all traffic on at once.

If you still need more after all that, each vendor has a form or a console option to request higher limits. Our overview of LLM API pricing compares the providers on cost.