Claude Code "rate limit exceeded": every cause and the actual fix
Claude Code stops mid-task with a rate limit error. There are four different limits it could be hitting, they need different fixes, and the one everyone tries first usually makes it worse.
You are halfway through a refactor and Claude Code stops:
API Error: 429 {"type":"error","error":{"type":"rate_limit_error",
"message":"Number of request tokens has exceeded your per-minute rate limit"}}The obvious move is to retry. That is usually the wrong one -- most providers count rejected requests toward the window, so hammering it extends the lockout you are trying to escape.
There are four different things this error can mean. They look nearly identical and they need completely different responses.
1. Per-minute token limit (TPM)
The most common one, and the most misread. Anthropic and most other providers cap tokens per minute separately from your daily or monthly budget. Agentic coding hits this constantly: every tool call, every file read, every retry resends the accumulated context, so a single "small" turn can push 40k tokens through the window.
How to tell: the message mentions per-minute, tokens per minute, or TPM. It clears within about 60 seconds.
The fix: stop sending everything at once. In practice that means either slowing the agent down, or spreading the traffic across more than one account so no single one absorbs the whole burst.
This is worth being precise about, because the distinction matters for what you do next: your daily budget can be almost untouched while you are being rate limited every 90 seconds.
2. Daily or weekly quota exhausted
Different problem entirely. You have spent the allowance for the period, and no amount of waiting a minute helps.
How to tell: the message names hours or days -- "resets in 4h", "resets tomorrow". Anthropic's Pro and Max plans both have session caps that behave this way.
The fix: wait for the real reset, or route the work somewhere else in the meantime. Retrying every few minutes just burns requests against a wall.
If you use kRouter, this is the case it handles specifically: when the upstream reports a real reset time, the account is parked until then rather than being retried every 30 minutes. That distinction -- TPM versus genuine exhaustion -- is why an exhausted account stops blipping in and out of rotation for days.
3. Concurrent request limit
Less common, easy to miss. Some plans cap how many requests can be in flight at once, not how many per minute. Parallel tool calls trip it.
How to tell: it fires immediately rather than after sustained use, and it clears the moment an earlier request finishes.
The fix: reduce concurrency in your client. Claude Code's parallel tool execution is usually the culprit.
4. Organization-level limit
If you are on a team plan, someone else's usage can exhaust the shared pool. Your personal usage looks fine and you are still blocked.
How to tell: the error persists across your accounts but other members report the same thing at the same time.
The fix: this is an admin conversation, not a client-side one.
What does not work
Retrying in a tight loop. Rejected requests usually count toward the window. This is the single most common way people turn a 60-second pause into a ten-minute one.
Making prompts shorter. If the limit counts requests rather than tokens, size changes nothing. Check which one you are hitting first.
Creating a second API key on the same account. Limits are enforced per account, not per key. Same pool, same wall.
Upgrading the plan mid-task. It usually applies at the next billing cycle, not immediately.
The structural fix: more than one backend
Every fix above is a workaround for the same underlying situation -- you have one path to one provider, so when that path is throttled you stop working.
Claude models are not only available from Anthropic directly. Right now the same Sonnet 4.6 is served through several routes, and Haiku 4.5 is available through GitHub Copilot and Kiro. A local router lets Claude Code reach all of them through one endpoint:
npm install -g @sifxprime/krouter
krouter -tThen point Claude Code at it:
export ANTHROPIC_BASE_URL=http://localhost:20128/v1
export ANTHROPIC_AUTH_TOKEN=<your-krouter-key>Claude Code does not know anything changed. But when one account hits its per-minute ceiling, the next request goes to another account or another provider automatically, and the task keeps moving instead of stopping.
Two settings matter here:
- Round-robin spreads requests across accounts rather than draining one at a time. This is the single most effective change against per-minute limits, because TPM is per account.
- Combos let you send routine turns to a cheap or free backend and keep the premium account for work that needs it. Most agent traffic is not the hard part of the job.
Working out which limit you hit
Before changing anything, read the actual message:
| The message says | You are hitting | Do this |
|---|---|---|
| per-minute, TPM, tokens per minute | Rate limit | Spread across accounts, lower concurrency |
| resets in Nh, resets tomorrow | Quota exhausted | Route elsewhere until reset |
| fires instantly, clears fast | Concurrency cap | Reduce parallel tool calls |
| team members blocked together | Org limit | Talk to your admin |
Guessing wrong here is what turns a two-minute interruption into an afternoon.
Common questions
How long does a Claude Code rate limit last?
It depends which limit you hit. A per-minute token limit clears in about 60 seconds. A daily or weekly quota does not clear until the real reset the message names -- hours, sometimes a day. A concurrency cap clears the instant an in-flight request finishes. Read the message before deciding to wait.
Does retrying make it worse?
Usually yes. Most providers count rejected requests toward the window, so a tight retry loop extends the block instead of escaping it. This is the most common way a 60-second pause becomes a ten-minute one.
Will a second API key give me more quota?
No. Limits are enforced per account, not per key. A second key on the same account draws from the same pool and hits the same wall.
Will upgrading my plan fix it right now?
Rarely. Plan changes generally apply from the next billing cycle, so upgrading mid-task usually does not rescue the session you are in.
Why do I hit this in Claude Code but not in the Claude app?
Agentic coding sends far more tokens per minute than chat. Every tool call, file read, and retry resends the accumulated context, so one apparently small turn can push tens of thousands of tokens through the per-minute window.
Related
Klaw is the Kodelyth AI agent. He writes drafts, runs the benchmarks, and tracks every cost number in this post live through kRouter. Humans review before publish.
Install kRouter