Anthropic rate limits explained: a deep dive for heavy users

You buy Claude Pro. You start a Cline session. Twenty minutes later you hit a wall:

"You have exceeded your usage limit. Please wait 5 hours."

You still have credits. You still have a working API key. So why?

Anthropic's rate-limit system is multi-dimensional, and most developers misunderstand at least one axis. Here is the full picture.

The four axes

Anthropic enforces limits along four separate dimensions, and you can hit any of them:

1. Requests per minute (RPM) -- pure call frequency
2. Input tokens per minute (ITPM) -- incoming token throughput
3. Output tokens per minute (OTPM) -- outgoing token throughput
4. Tier-bound usage windows -- the "5-hour quota" on Pro plans, and weekly bundles on Max

If a single one of these caps trips, the entire request gets a 429 -- even if the other three have headroom.

The tier ladder

You auto-promote as you spend more. New accounts always start at Tier 1, which is why your first day with Claude often feels frustrating.

The Pro-plan trap

Pro subscribers (claude.ai + Claude Code CLI) have a separate quota system from the API. The 5-hour usage window on Pro is not based on raw tokens -- it is based on Anthropic's opaque "fair-share" calculation that considers your traffic shape, peer comparison, and model complexity.

Heavy Cline users on Pro often hit the wall in 90 minutes because Cline's agentic loops register as bursty traffic.

The TPM vs daily quota trap (v0.5.49)

Before v0.5.49, kRouter treated all Anthropic 429s the same way: back off for the retry-after duration and try again. This caused a subtle bug.

Anthropic actually returns two different kinds of rate limits:

• TPM (tokens per minute) -- resets every 60 seconds. Short cooldown.
• Daily quota -- resets at midnight UTC. Long cooldown.

If you hit a daily quota limit, backing off for 60 seconds just wastes time -- you will get another 429. v0.5.49 added TPM vs daily quota disambiguation: kRouter now parses the retry-after header and the error body to distinguish between the two. Daily quota hits mark the account as exhausted until the next UTC midnight. TPM hits get the standard short cooldown.

This means kRouter stops hammering a daily-exhausted account and immediately falls through to the next provider in your combo chain.

The permanent ban flag (v0.5.47)

Some Anthropic accounts get permanently suspended -- usually for terms-of-service violations or payment disputes. Before v0.5.47, kRouter would keep trying these accounts on every request, eating latency each time.

v0.5.47 wired the permanent ban flag: when kRouter detects a 403 with specific ban indicators, it marks the account as permanently disabled. No more wasted requests. The account stays disabled until you manually re-enable it in the dashboard.

The Test Connection lock fix (v0.5.44)

The "Test Connection" button in kRouter's dashboard used to acquire a SQLite lock that was not released on timeout. If the test failed (which it does when the account is rate-limited), the lock would hold for 30 seconds, blocking all routing decisions.

v0.5.44 fixed this with proper lock clearing on timeout. A failed test connection no longer freezes the router.

How kRouter's atomic backoff prevents retry storms

The naive approach to rate limits: catch the 429, wait, retry. The problem: if you have 10 concurrent requests and they all get 429s at the same instant, they all wait the same duration and all retry at the same instant -- creating a retry storm that triggers another 429.

kRouter uses atomic backoff with SQLite transactions:

1. When a 429 arrives, kRouter opens a SQLite transaction.
2. Inside the transaction, it reads the current cooldown state for that account.
3. If no cooldown is set, it writes the new cooldown timestamp and commits.
4. If a cooldown is already set (another concurrent request beat us), it skips the write and falls through to the next provider immediately.

This means exactly one request sets the cooldown. All other concurrent requests see the cooldown and route elsewhere instantly. No retry storm. No wasted time.

How to engineer around it

Option 1: Pay for a higher tier. Spend $40 in a week to unlock Tier 2, which is 20x the RPM of Tier 1. Fast but expensive.

Option 2: Multi-account routing through kRouter. Configure two or three separate Anthropic accounts as providers in kRouter, and let it round-robin between them:

1. anthropic-acct-a/claude-sonnet
2. anthropic-acct-b/claude-sonnet
3. anthropic-acct-c/claude-sonnet

Each account has its own independent rate limit. You get 3x the headroom for no extra cost.

Option 3: Smart fallback to non-Anthropic. Configure GLM-5.1 or DeepSeek as a fallback. When Anthropic rate-limits you, requests silently switch to a comparable model:

1. anthropic/claude-sonnet
2. glm/glm-5.1        # 5x cheaper, comparable quality

Option 4: Let atomic backoff do its job. With kRouter's rate-limit handling, you do not need to think about retry logic. The router respects the exact reset window, disambiguates TPM from daily limits, flags permanent bans, and routes around exhausted accounts. Your IDE keeps working.

The bottom line

You cannot fight Anthropic's rate limits with retries. You engineer around them by spreading load across accounts and falling back to cheaper providers. kRouter automates both. See the full routing configuration on /install and the rate-limit changelog entries on /changelog.

npm install -g @sifxprime/krouter