Claude Code Rate Limit: What "Usage Limit Reached" Means and 9 Ways to Make Your Plan Last

If Claude Code just stopped with a line about a limit, the first thing to know is that there are four different messages and they do not mean the same thing. Three belong to a subscription and are about how much you have used. One belongs to the API and is about how fast you are sending. The fix depends on which you hit, so start there.

Nothing in this article lifts a limit. Limits are enforced on Anthropic's side and no setting, tool or trick changes them. What you can change is how quickly you reach them, and that turns out to be a lot.

Everything below was checked against Anthropic's Claude Code documentation on 18 September 2026.

Which limit did you hit?

Message	What it is	Does /model help?
You've hit your session limit	Subscription. The rolling five-hour window.	No, shared across all models
You've hit your weekly limit	Subscription. The weekly window.	No, shared across all models
You've hit your Opus limit / Sonnet limit	Subscription. A cap on one model family.	Yes: switch to a model outside that family
Request rejected (429) / Server is temporarily limiting requests	API key, Console or cloud provider. Too many requests right now.	No; Claude Code retries by itself

There is a fifth family that looks similar and is not a usage limit at all: You've hit your monthly spend limit and its variants. That is a cap you or your admin set on usage credits, and the fix is to raise it or wait for the plan reset the message names.

And one false alarm: a context or auto-compact warning is not a limit. It means the conversation has grown close to the window where Claude Code summarises old history. It is, however, a strong hint about why you are hitting limits, which we will get to.

How the subscription windows work

On Pro, Max, Team and Enterprise, your Claude Code usage draws on an allowance that resets on a rolling five-hour window and a weekly window. Three details matter:

• It is shared. Claude Code, Claude chat and Cowork draw from the same allowance. A long research afternoon in the app is an afternoon your terminal does not get.
• The size depends on the plan, and Anthropic does not publish it as a token count. Max 5x and 20x are multiples of Pro; Team and Enterprise depend on the seat tier.
• Coding costs more than chatting. Anthropic's own guidance to admins is to budget more for a coding seat than a chat seat, because each Claude Code turn carries file contents, tool calls and multi-step reasoning. In their words, one debugging session can consume more than a day of chat.

What Claude Code does when you hit it

Since v2.1.234, Claude Code does not just stop. In an interactive session on a subscription it waits and shows:

Usage limit reached · continuing automatically at 3:45pm · esc to cancel

Leave the session open and it picks the task up where it stopped after the reset, with a fixed prompt rather than by resending your last message. If your computer slept for more than about half an hour it asks you to press Enter. It still asks for permissions as usual, so a task can sit on a permission prompt while you are away, and if it hits the limit again it re-arms at most twice in a row before giving up.

Press Esc or run /rate-limit-options and choose Don't continue automatically to cancel. The same menu lets you start the wait again.

Your real choices at that moment:

1. Wait. Free. The message tells you when.
2. Switch model, if and only if the message named a model.
3. Turn on usage credits with /usage-credits. You keep working past the allowance at standard API rates, up to a monthly spend limit you set. Set one.
4. Upgrade the plan with /upgrade.
5. Use less, so that next time option 1 comes later. The rest of this article.

First, find out what is eating your plan

Run /usage. On a subscription it shows your plan bars and, more usefully, a breakdown of what counted against them:

• Attribution: recent usage by skill, subagent, plugin and individual MCP server, as percentages.
• Behaviour flags: anything that accounts for 10% or more of recent usage, such as long context or cache misses, each with a tip.
• Loops: the heaviest /loop and scheduled tasks, with how often they fire and tokens per run.

Press d or w to switch between the last 24 hours and the last 7 days. The figures come from session history on this machine, so usage from other devices or from claude.ai is not in them. /context shows what is filling the current conversation. Two minutes with these two commands usually names the culprit.

Why a long session burns your plan faster than you think

This is the part that surprises people. Claude Code sends your full conversation with every request, and every time Claude uses a tool it sends another request carrying that batch of tool results. Prompt caching makes the re-read cheap, but not free, so a one-line question at 5pm in a session you opened at 9am still draws usage for the whole day's conversation.

Anthropic's docs list the usual causes of a session that drains a plan while you feel idle:

• Long context. Everything above.
• Cache misses. The cache lifetime is an hour on a subscription. Come back from a long lunch and your next message reprocesses the entire context at full price. Once you are drawing on usage credits, the lifetime drops to five minutes unless you set it yourself.
• Scheduled tasks and loops firing on their interval in an idle session, sending the full context each time.
• Goal check-ins and cross-session messages starting new turns while you are away.
• Agent teammates, each with its own context window, consuming until they exit. Agent teams in plan mode use roughly seven times the tokens of a standard session.
• Compaction itself. /compact has to read the conversation it summarises, so compacting a huge context is a huge request. /clear costs nothing.

Nine ways to make the same plan last

Ordered by how much they usually save for how little effort. The first seven are Anthropic's own advice; none of them needs a tool.

1. /clear between unrelated tasks. The highest-impact habit on the list. Stale context is paid for again on every message. /rename first so you can /resume later.

2. Match the model to the job. Sonnet handles most coding work and draws less than Opus. Keep Opus for architecture and multi-step reasoning, switch with /model, and set model: haiku on simple subagents. Anthropic names "Opus left as the default" as one of the two usual causes of unexpected spend.

3. Do not let the cache go cold mid-task. If you are stepping away for more than an hour, finish the thought first or /clear when you return. On Pro and Max, Claude Code offers to resume a large session from a summary after a long break: take it.

4. Find your loops. /usage lists them. A /loop checking a deploy every five minutes in a session with a day of history is an expensive clock.

5. Send noisy work to subagents and hooks. Test runs, log files and documentation fetches stay in the subagent's context and only a summary comes back. A hook can grep a 10,000-line log down to the error lines before Claude ever sees it.

6. Shrink what loads at start. CLAUDE.md is in context for every request; aim for under 200 lines and move workflow-specific instructions into skills, which load only when used. Run /mcp and disable servers you are not using; prefer CLI tools such as gh where one exists. Every MCP server you keep is also a door for indirect prompt injection, so fewer is safer as well as cheaper.

7. Plan first, be specific, stop early. Plan mode (Shift+Tab) before a big change prevents paying for the wrong approach. "Add input validation to the login function in auth.ts" costs a fraction of "improve this codebase". When Claude heads the wrong way, press Escape at once rather than letting it finish. Lower /effort for simple tasks: thinking tokens are billed as output.

8. Keep the chat app for chat. The allowance is shared. If your weekly limit keeps arriving on Thursday, look at what you do in claude.ai on Monday.

9. Send fewer tokens per turn. Habits 1 to 7 reduce what you put into context. The remaining waste is structural: Claude Code re-sends files it has already read, tool schemas, terminal colour codes, duplicate log lines and its own scaffolding on every turn. You cannot fix that with a habit, but a gateway can strip it before it reaches Anthropic.

That is what we use Constellation Gate AI for (our review). It sits "between your agent and the model" and removes "token waste", "inefficient tool calls and duplicate content", keeping the first copy of everything so "the model receives the same content and returns the same answer". Gate's published figure is 20% or more fewer tokens per request on agent workloads, 23% in its Rocket Resume case study (vendor numbers, checked 18 September 2026). It works with your subscription untouched, because its key travels in a custom header rather than in the API credential slot; the proxy guide has the exact setting.

Let us be exact about the claim, because this is where the internet gets sloppy. Gate does not raise, lift or alter your limit. The limit is the same number it was yesterday. If fewer tokens reach Anthropic per turn, the same allowance should cover more turns. How much of Gate's 20% shows up as extra runway on a Max plan depends on how Anthropic meters cached versus uncached tokens against the allowance, which is not published. We are measuring it on our own sessions and will put the number here when we have it rather than guess.

Free plan: 20,000 recorded requests a month with basic compression. The full compression is Pro, $20 per user per month. Affiliate link: we earn a commission on Pro seats, which pays for the testing and has no say in the verdict.

If you are on an API key instead

Then you have no session or weekly limit; you have a bill and a rate. Request rejected (429) means the API is limiting how fast you send, and Claude Code retries by itself. If you see it often, your organisation's tokens-per-minute allocation is too small for how many people share it. Anthropic's guidance for small teams of one to five is 200k to 300k TPM and 5 to 7 requests per minute per user, falling as the team grows because fewer people work at the same moment.

Everything in the nine habits applies twice over, because here each token is money rather than allowance. The Session block at the top of /usage shows the running cost, and a gateway with a spend cap turns "I hope the agent did not loop overnight" into a number it cannot exceed.

Your first fifteen minutes

Following the rule we apply to every tool: real input, one output, hard stop.

1. Minutes 0 to 3. Run /usage, press w. Write down the top two lines of attribution and any behaviour flag.
2. Minutes 3 to 6. Run /context in the session you have had open longest. Note what is filling it.
3. Minutes 6 to 12. Fix the one thing those two screens pointed at: /clear the stale session, change the default model in /config, stop the loop, disable the MCP server you forgot about, or cut CLAUDE.md.
4. Minutes 12 to 15. Write today's date and the weekly bar's percentage on a sticky note.

Output: one fix and one baseline. Next week the same bar tells you whether the habit held. If you want the structural saving as well, route Claude Code through Gate (or OpenCode or Codex) and put the dashboard's tokens-saved number next to it.

The limit is a fixed size. Most of what fills it is not your work; it is yesterday's context, a model bigger than the task, and the same file sent for the fortieth time. Clear those out and the plan you already pay for gets noticeably larger without changing at all.