Claude API Rate Limits: 429 and 529 Errors Explained

Claude API rate limits by usage tier: RPM, ITPM and OTPM per model, spend caps, ramp limits, the rate limit headers, and how to handle 429 and 529 errors.

Josh Pocock8 min read

API rate limits cover

AI Coding

API rate limits

Stride AI Academy

A Claude API rate limit caps how many requests and tokens your organization can send per minute. Limits are set per model and per usage tier, measured in requests per minute (RPM), input tokens per minute (ITPM) and output tokens per minute (OTPM). Go over one and the API returns a 429 rate_limit_error. A 529 overloaded_error is different: it means the API itself is busy, not that you did anything wrong.

This guide covers the usage tiers and spend caps, the per-model limits, how caching raises your real throughput, ramp limits, the response headers, and retry code for 429 and 529 errors. Everything here was checked against Anthropic's rate limits and errors docs and the Python SDK docs on November 22, 2026. Limits change, so check the Console for your own numbers.

If you hit limits in the Claude apps or in Claude Code on a subscription, that's a plan usage limit, not an API rate limit. Our guide to Claude usage limits covers those.

Two kinds of Claude API limits

The rate limits docs describe two:

  1. Spend limits cap what your organization can spend on the API each month.
  2. Rate limits cap how many requests and tokens you can send over a period of time.

Both are set at the organization level, and you can set lower limits for individual workspaces. You can see your tier and current limits on the Rate limits page in the Claude Console.

Claude API usage tiers and spend caps

Anthropic places your organization on a tier automatically, based on usage history and account standing. You move up over time as you use the API. New organizations can start on an Evaluation tier with limits below the standard ones while account history builds.

Each standard tier has a monthly spend cap:

Usage tier Monthly spend cap
Start $500
Build $1,000
Scale $200,000
Custom No cap, arranged with your account team

When you reach your tier's cap, API usage pauses until 00:00 UTC on the first day of the next month. Requests return a 429 with error_code set to enforced_spend_limit_reached and no retry-after header, so retrying won't help. Request a tier increase on the Rate limits page instead.

You can also set your own lower limit under Spend limits. When you hit a limit you set, requests return a 400 invalid_request_error, not a 429.

Rate limits per model: RPM, ITPM and OTPM

Here are the standard limits for the current models on November 22, 2026, from the rate limits docs:

Model Start tier (RPM / ITPM / OTPM) Build tier Scale tier
Fable 5.2 4,000 / 500,000 / 100,000 10,000 / 1,500,000 / 300,000 13,000 / 4,000,000 / 800,000
Opus 5.5 4,000 / 2,000,000 / 400,000 10,000 / 5,000,000 / 1,000,000 13,000 / 10,000,000 / 2,000,000
Sonnet 5.5 4,000 / 2,000,000 / 400,000 10,000 / 5,000,000 / 1,000,000 13,000 / 10,000,000 / 2,000,000
Haiku 5.5 4,000 / 2,000,000 / 400,000 10,000 / 5,000,000 / 1,000,000 13,000 / 10,000,000 / 2,000,000
Opus 4.x, Sonnet 4.x, Haiku 4.5, Opus 5, Sonnet 5 1,000 / 2,000,000 / 400,000 5,000 / 5,000,000 / 1,000,000 10,000 / 10,000,000 / 2,000,000

A few details matter:

  • Some models share a bucket. Opus 4.8, 4.7, 4.6 and 4.5 share one combined limit. Sonnet 4.6 and 4.5 share another. Opus 5.5, Opus 5, Sonnet 5.5 and Sonnet 5 each have their own.
  • Limits can bite over shorter windows. The docs say 60 RPM might be enforced as 1 request per second, so a short burst can trigger a 429 even if your minute total is fine.
  • Capacity refills continuously. The API uses a token bucket, so you don't wait for a fixed reset.
  • max_tokens doesn't count against OTPM. Only tokens actually generated do.

Caching raises your real ITPM

The default ITPM limit counts only uncached input tokens. Tokens read from the prompt cache don't count. The docs give this example: with a 2,000,000 ITPM limit and an 80 percent cache hit rate, you could process 10,000,000 total input tokens a minute. Tokens written to the cache do count. So caching your system prompt, tool definitions and long documents is a cheap way to get more throughput.

Ramp limits

A sharp jump in usage can cause a 429 even when you're under your rate limit. The ramp limit is your recent peak per-minute usage plus a step, and it rises as your traffic grows. On the Messages API these errors carry error_code: "ramp_limit_exceeded". The docs say to wait at least the retry-after seconds, hold traffic where requests succeed, then grow by at most one step per minute. Start bulk jobs slowly rather than all at once.

The Message Batches API and fast mode have their own separate limits, and batches aren't subject to ramp limits.

Claude API rate limit headers

Every response tells you where you stand. The main headers:

Header What it tells you
retry-after Seconds to wait before retrying. Not every 429 has it
anthropic-ratelimit-requests-remaining Requests left before you're limited
anthropic-ratelimit-input-tokens-remaining Input tokens left, rounded to the nearest thousand
anthropic-ratelimit-output-tokens-remaining Output tokens left, rounded to the nearest thousand
anthropic-ratelimit-tokens-reset When the token limit fully refills, in RFC 3339 format

Each has matching -limit and -reset versions. The anthropic-ratelimit-tokens-* headers show whichever limit is most restrictive right now, such as a workspace limit.

To read them in Python, use the raw response:

python
from anthropic import Anthropic

client = Anthropic()

response = client.messages.with_raw_response.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)

print(response.headers.get("anthropic-ratelimit-requests-remaining"))
print(response.headers.get("anthropic-ratelimit-output-tokens-remaining"))
message = response.parse()

Claude API 429 vs 529

429 rate_limit_error 529 overloaded_error
Cause Your organization hit a rate limit, a ramp limit, or its spend cap The API is temporarily overloaded across all users
Your fault? Yes, in the sense that it's your traffic No
What to do Wait for retry-after, slow down, cache more, or request a higher tier Retry with backoff, or try again later

The errors docs list the others you might see: 400 bad request, 401 key problems, 402 billing, 403 permissions, 404 not found, 413 request too large, 500 internal error and 504 timeout. Every error response includes a request_id, so log it and include it if you contact support.

If 529s pile up across everything, check whether Claude is down before you debug your own code.

How to handle Claude API rate limit errors

Start with what the SDK already does. The official SDKs retry connection errors, 429s and 5xx errors twice by default with exponential backoff, and they honor retry-after when it's present. You can raise or disable that with max_retries:

python
from anthropic import Anthropic

client = Anthropic(max_retries=5)  # default is 2

If you want your own loop on top, catch the SDK's typed errors. The Python SDK raises RateLimitError for a 429 and InternalServerError for status codes of 500 and up, which includes a 529:

python
import random
import time

import anthropic

client = anthropic.Anthropic(max_retries=0)


def ask(prompt, attempts=6):
    for attempt in range(attempts):
        try:
            return client.messages.create(
                model="claude-sonnet-5-5",
                max_tokens=1024,
                messages=[{"role": "user", "content": prompt}],
            )
        except (anthropic.RateLimitError, anthropic.InternalServerError) as e:
            if attempt == attempts - 1:
                raise
            retry_after = e.response.headers.get("retry-after")
            if isinstance(e, anthropic.RateLimitError) and retry_after is None:
                raise  # likely a spend cap: retrying will not help
            wait = float(retry_after) if retry_after else min(2 ** attempt, 60)
            time.sleep(wait + random.random())

The check for a missing retry-after on a 429 follows the docs' note that the spend-cap 429 never carries one. Treat it as a stop, not a pause.

Beyond retries, these cut rate limit errors at the source:

  • Cache repeated input. Cached reads don't count toward ITPM.
  • Spread requests out. Even pacing beats bursts, and it keeps you under ramp limits.
  • Use the Batch API for work that can wait. It has separate limits and, per the pricing docs, costs 50% less.
  • Pick the right model. Opus 5.5, Sonnet 5.5 and Haiku 5.5 have higher RPM than the older models on every tier.
  • Watch the Usage page. Its rate limit charts show your peak per-minute usage against your limit and your cache rate.

How to raise your Claude API limits

Use Request tier increase on the Rate limits page in the Console. For limits above the Scale tier, contact sales from the same page. Your tier also rises on its own as you build usage history.

To protect one project from another, set lower limits on individual workspaces. Organization-wide limits always apply on top. For your first requests and key setup, see the Claude API tutorial and how to get a Claude API key.

Learn to build with Claude Code

The Stride AI Coding course teaches Claude Code and Codex for building client tools, written for non-programmers, with weekly live calls to review your builds. Many of the free guides cover Claude Code setups, and the Claude Code community is where members compare theirs.

FAQ

What is the Claude API rate limit?

It depends on your usage tier and the model. On the Start tier on November 22, 2026, Opus 5.5, Sonnet 5.5 and Haiku 5.5 each allow 4,000 requests, 2,000,000 input tokens and 400,000 output tokens per minute. Your own limits are on the Rate limits page in the Claude Console.

What does a Claude API 429 error mean?

Your organization went over a rate limit, a ramp limit or its monthly spend cap. Wait the number of seconds in the retry-after header and slow down. A 429 with no retry-after is usually the spend cap, which retrying won't fix.

What is a Claude 529 error?

A 529 overloaded_error means the API is temporarily overloaded across all users. It isn't caused by your traffic. Retry with exponential backoff, and check the status page if it continues.

Do cached tokens count toward Claude API rate limits?

No. Tokens read from the prompt cache don't count toward your input tokens per minute limit. Uncached input and tokens written to the cache do, which makes caching a direct way to raise throughput.

How do I increase my Claude API rate limits?

Click Request tier increase on the Rate limits page in the Claude Console. Anthropic also moves organizations up tiers automatically as usage history grows. For limits above Scale, contact sales.

Build it with us in Stride AI Academy

Four courses, weekly live calls and a community of people building and selling AI. Or start with 320+ free guides.

Keep reading

Free AI Coding guides