Guide

Anthropic API Rate Limits & Usage Tiers Explained (2026)

2026-06-29 8

TL;DR — Anthropic meters the Claude API on three axes at once: requests per minute (RPM), input tokens per minute (ITPM), and output tokens per minute (OTPM). Your ceilings are set by a usage tier (1–4) that you climb automatically as your cumulative spend grows. Hit any ceiling and you get an HTTP 429 with a retry-after header — back off and retry. A useful 2026 detail: on most models cached input tokens don't count toward ITPM, so good prompt caching effectively raises your limit.

The three limits that matter

Every Claude API key is rate-limited on three dimensions simultaneously. You hit a 429 when you cross any one of them, not all three:

Limit What it counts Bites when…
RPM — requests/minute Number of API calls You fire many small requests (parallel agents, batched UI calls)
ITPM — input tokens/minute Input tokens sent You send large prompts (big repos, long histories) at high frequency
OTPM — output tokens/minute Output tokens generated You generate long outputs at high frequency

Because the three are independent, the fix depends on which one you're hitting. A swarm of tiny requests trips RPM; a few giant-context requests trip ITPM. Read the error — Anthropic returns headers (anthropic-ratelimit-*) telling you which limit and how much remains.

Usage tiers: how the ceilings are set

Your limits aren't fixed — they scale with a usage tier, and the tier is driven by cumulative spend, not monthly spend:

Tier Roughly unlocked at (cumulative) Character
Tier 1 ~$5 Entry limits — fine for prototyping, tight for production
Tier 2 ~$50 Comfortable for small apps
Tier 3 ~$100 Real production headroom
Tier 4 ~$400 Maximum self-service limits

Beyond Tier 4, higher ceilings require a sales conversation. You advance automatically once cumulative credits cross the threshold — so if you know you'll need Tier 3 limits, front-loading credit gets you there sooner. (Exact per-tier RPM/ITPM/OTPM numbers shift over time and vary by model — always confirm yours on the official Limits page; the structure above is what's stable.)

The cache detail that effectively raises your limit

Here's the lever most people miss: on most Claude models, only uncached input tokens count toward ITPM. Cached reads are ignored by the rate limiter.

So if you run an agentic workload at an 80% cache-hit rate against a 2M ITPM ceiling, you can actually push ~10M input tokens/minute through — because 8M of them are cached reads the limiter doesn't count. Prompt caching isn't just a cost saver (cached reads bill at ~10% of normal — see Claude pay-as-you-go pricing); it's also a throughput multiplier. For how to structure prompts so they actually cache, see Claude Code token cost and context management.

Handling a 429 cleanly

A 429 is throttling, not a broken key. The right response is mechanical:

  1. Read retry-after — it tells you how many seconds to wait. Honor it.
  2. Exponential backoff with jitter — official SDKs already retry 429/5xx automatically (default 2 retries); only hand-roll if you need more.
  3. Check which limit you hit via the anthropic-ratelimit-* headers, and shape your traffic accordingly — fewer/larger requests for RPM, more caching for ITPM.

For a Claude Code-specific walkthrough, see fixing Claude Code's 429 rate limit error.

Where a pooled relay changes the picture

Per-tier limits exist because you're one account talking to one endpoint. A relay sits in front of a pool of upstream capacity, which changes two things:

  • No tier application to manage. You don't climb a spend ladder or file limit-increase requests — you use the relay's pooled headroom from day one.
  • Failover across upstreams. When one upstream throttles, a multi-upstream relay can route around it, so a single 429 upstream doesn't stall you.

LLM API works this way: one sk- key against pooled, high-availability capacity, with both the Anthropic and OpenAI protocols on the same base URL. The honest trade is the one every relay carries — you're routing through a third party (more in what a Claude-compatible API relay is). Live rates are on the pricing page; grab a key at /register.

FAQ

What's the difference between RPM, ITPM, and OTPM? RPM caps how many requests you make per minute; ITPM caps input tokens per minute; OTPM caps output tokens per minute. You're limited by whichever you hit first — they're enforced independently.

How do Anthropic API usage tiers work? There are four tiers. Your tier is set by cumulative spend (~$5 / ~$50 / ~$100 / ~$400), and you advance automatically as credits accumulate. Higher tiers raise your RPM/ITPM/OTPM ceilings; beyond Tier 4 you contact sales.

Do cached tokens count toward my rate limit? On most Claude models, no — only uncached input tokens count toward ITPM. A high cache-hit rate effectively multiplies your usable throughput.

Why am I getting a 429 if my key is valid? A 429 means you crossed a rate limit (RPM, ITPM, or OTPM), not that your key is bad. Read the retry-after header, back off, and check the anthropic-ratelimit-* headers to see which limit you hit.

How do I get higher Claude API rate limits? Either climb the usage tiers by spending more (cumulative), request an increase on the Limits page, or route through a pooled relay that gives you shared high-availability capacity without per-tier application.

llmapi.pro is an independent, Claude-compatible API relay; we are not affiliated with Anthropic. Claude is used for identification only. Tier thresholds and per-tier limits are summarized for reference — confirm current values on Anthropic's official Limits page.

Share this article

Start using LLM API

Free tier available. One-line configuration for Claude Code.

Get Started Free