Claude Code Usage Limits: What Runs Out, and When It Comes Back

Claude Code stops mid-task and prints one line with a time in it. That line is the only place the tool tells you which ceiling you just reached, and it can name three different ones.

Claude Code enforces three separate ceilings: a rolling session limit, a weekly limit, and a model-specific Opus limit. The message it prints names which one and when it resets. As of 2026-08-16, Anthropic publishes no number for any of the three.

Anthropic’s own support articles, product documentation and announcement post are where every mechanic below comes from, each named where its claim appears and each read on 2026-08-16. No account was opened for it, and no limit was hit on purpose.

The last sentence of that answer is the awkward one, because the figures in circulation are very specific. One ranking guide prices a Pro window at about 44,000 tokens. Another prices it at 10 to 45 prompts and a Pro week at 40 to 80 Sonnet hours. A third counts a Max 20x window in messages. Every one of those numbers has a source, and none of the sources is Anthropic. One name still needs untangling before any of that: “Claude Code rate limits” covers two different systems, the subscription window and the organization-level rate limit on the developer platform, which behaves nothing like it. The FAQ below separates them.

What actually runs out when Claude Code stops

The Claude Code errors reference, checked 2026-08-16, collects this whole class of stop under one heading and prints three messages under it:

You've hit your session limit · resets 3:45pm
You've hit your weekly limit · resets Mon 12:00am
You've hit your Opus limit · resets 3:45pm

The reset time sits inside the message. That is the most useful property these three lines have, and it is why the first move is to read the line rather than go looking for a dashboard.

Two other stops look like usage limits in a terminal and are not. The Claude Code cost documentation separates them for admins under the heading “When a developer asks about a limit”: a context or auto-compact warning is not a usage limit, and a spend cap on a self-hosted gateway is a ceiling the operator set rather than an allowance on your plan. A third one does too: a line that starts API Error: 500 is Anthropic’s own side failing rather than any ceiling on your account, and what a 500 leaves behind is a different question from when your window resets.

The first two rows are also why a heavy morning in the Claude app can end your afternoon in the terminal. Anthropic’s help center article on usage and length limits, dated 2026-07-13, states the shared-pool rule in one sentence: “Note that your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit.”

What the table cannot show you is that a clock is only one way to ration a coding tool. A product that meters a monthly balance instead stops you differently, and that difference shows up on the day you are trying to finish something rather than on a pricing page.

The five-hour window, the weekly one, and when each resets

Anthropic calls the short window rolling. The cost documentation describes a seat allowance that “resets on a rolling five-hour window and a weekly window”, and third-party guides usually write the same thing as the 5 hour limit. What none of the pages checked here states is what starts that window. Two ranking guides say it begins with your first message, and that claim appears on no Anthropic page in this set.

The weekly message is more concrete about its reset than the session one, because it prints a weekday: resets Mon 12:00am. How that weekday is chosen for an account is not documented on the errors reference either.

On claude.ai the same two windows appear as bars. Anthropic’s usage limit best practices article, dated 2026-06-02, describes it plainly: “If you’re using a Pro, Max, Team, or seat-based Enterprise plan, you can navigate to Settings > Usage to view progress bars showing how much of your five-hour session and weekly usage limits you’ve consumed.” That page also says the weekly section shows resets “for Opus only and all other models”, which is the clearest published confirmation that Opus is metered on its own weekly track.

What no page gives is a size. Six Anthropic pages were read in full on 2026-08-16 looking for one: four support articles and two product documentation pages. Article 11647753 lists the factors that affect usage and gives no counts. Article 11145838 covers what happens when you reach a limit and gives no counts. Article 14552983, “Models, usage, and limits in Claude Code”, dated 2026-04-15, describes a pool “reset on a rolling window” and gives no counts. The errors reference and the cost documentation give none either.

As of 2026-08-16, Anthropic’s Claude Code and plan documentation does not publish a numeric per-window cap for any plan.

That is a statement about six pages on a specific date, not a claim about what Anthropic has ever said. It is also the single fact that makes every table further down this page readable.

Do Claude Code limits differ by plan?

Plans differ in which windows apply and where the allowance is shared, and Anthropic documents both. They do not differ in any published number, because there is no published number to differ by. Free is the one clean exception, and only because it has no Claude Code access at all.

The seat rows come from the cost documentation, which states that on Teams and Enterprise plans each member’s usage draws from a per-seat allowance on those two windows shared with Claude chat and Cowork. The usage-based row comes from the best practices article, which says such an organization has no specific usage limits and is charged on consumption.

Claude Code limits on a Max plan are those same two windows, which is why the Max row is the whole problem in one line. The plans are named for a multiple. The thing being multiplied is not published.

Model-specific limits, and why Opus stops before Sonnet does

The errors reference is direct about which stop a model switch can rescue you from: “The session and weekly limits are shared across all models, so switching models doesn’t restore access. The Opus limit applies only to Opus requests, so switching to another model with /model keeps you working.”

So one of the three messages is worth trying something about and two are not. If the word Opus is in the line, /model is the answer and it takes seconds. If it is not, the reset time is the answer.

Why Opus runs out first is documented in relative terms only. Article 14552983 says Opus offers deeper reasoning for harder problems, uses meaningfully more of your quota, and is worth switching to when you need it rather than leaving on. The cost documentation gives the same advice from the other end, recommending Sonnet for agent-team teammates because it balances capability and cost.

The exchange rate is the part nobody publishes. No page in this set says an Opus request counts for some multiple of a Sonnet request against your window. Anyone printing one is estimating.

Where every Claude Code token limit number you have read came from

Three of the pages ranking for this family print a figure for what a window allows. They disagree on the figure, and, more usefully, they disagree on the unit. faros.ai counts allocated tokens. truefoundry counts prompts per window and Sonnet or Opus hours per week. claudefa.st counts messages, describing a Max 20x account that “used to cap around 900 messages per 5-hour window” and putting the doubled figure at “~1,800”. Its page does link Anthropic’s announcement, for the change itself; the announcement contains no message counts.

None of the three is linked here. truefoundry sells an LLM gateway, which puts it under the standing rule against sending readers to a company selling into the same problem. faros.ai and claudefa.st are unlinked for an editorial reason: they are the pages whose figures this section is checking, and linking them hands you the number that was just flagged.

AxonBuild’s read of all three on 2026-08-16 found the same shape every time: the table is stated and the method is not. Three ranking pages, three incompatible units, and not one figure that survives a trace back to a page Anthropic publishes.

That does not make any of them wrong. A number derived from watching a real account run out is worth something. It does make them unverifiable, which is why two people on one plan can read two guides and come away with two incompatible models of the same product.

What changed in 2026, and why your limits feel different

One dated change explains most of the “these got worse” and “these got better” traffic at once. Anthropic’s announcement of higher usage limits, dated 2026-05-06, lists three moves and two of them land on Claude Code: “First, we’re doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans.” and “Second, we’re removing the peak hours limit reduction on Claude Code for Pro and Max accounts.”

The second sentence answers a question the first does not. It establishes that a peak-hours reduction existed, which is the mechanism behind a whole category of complaint about a window that drained faster on a weekday morning than a weekend one.

claudefa.st reads the same announcement narrowly, and correctly: the change moves the five-hour rate limits and removes the peak-hours reduction, and says nothing at all about the weekly cap. Its write-up concludes the weekly ceiling stayed the size it was.

A doubling, though, is a claim about a ratio. Without a published base, a reader outside Anthropic cannot check what doubled, cannot confirm it landed on their account, and cannot tell a real increase from a quiet week. Every “limits doubled” figure in circulation is somebody’s estimate of the old number multiplied by two.

The staleness runs the other way too. faros.ai’s page, in its 2026-06-26 update, still describes the March 2026 peak-hours reduction and makes no mention of the 2026-05-06 announcement removing it. Anyone reading that page today is reading a description of a behavior Anthropic said it withdrew seven weeks earlier.

How to see where you actually stand

/usage is the answer inside Claude Code, and on a Pro, Max, Team, or Enterprise plan it shows more than bars. The cost documentation describes two blocks worth knowing about. Attribution breaks recent usage down by skill, subagent, plugin, and individual MCP server, each as a percentage of the total. Behavior flags mark any behavior accounting for 10% or more of recent usage, such as long context or cache misses. Press d or w to switch between the last 24 hours and the last 7 days.

Two caveats matter more than the figures do. They are approximate, and they are computed from local session history on that machine, so a second laptop and a busy morning in the browser are both invisible here while both spend the same allowance.

That breakdown is also the only place the running cost of a connected MCP server appears as a share of a window rather than as a line in a config file. A server taking a tenth of your recent usage that you cannot remember the last useful answer from is a decision about which servers are worth keeping connected, not a limits problem.

When the usage request itself gets rate limited, /usage falls back to the last bars it loaded on that machine within the past 60 minutes and labels them Showing last-known usage with the age of the data. Press r to retry. Without a snapshot from the past hour, it reports the endpoint as rate limited and offers the same retry.

Three surfaces outside the command show the same allowance. The errors reference names two: add the rate_limits fields to a custom status line, or click the usage ring next to the model picker in the Desktop app. The third is Settings then Usage on claude.ai, the only one that covers every device and the chat product at once.

Why a quiet session still eats your window

The most common complaint about these limits is that they arrive after what felt like an hour of light work. The cost documentation explains that under “Why usage climbs in a long session”, and the mechanism is not subtle: Claude Code sends your full conversation with every request, and each tool use sends another request carrying that batch of results. With prompt caching that history is re-read at the cached rate, so, in Anthropic’s words, “a one-line question in a session that has been open all day still draws usage for the whole conversation.”

Five other things spend the allowance without you typing anything, all from the same page:

Cache misses come first. The prompt cache lifetime is an hour on a subscription and drops to five minutes once you are drawing on usage credits, and it is five minutes by default on an API key or cloud provider. Your first message after a break longer than that reprocesses the full context. Setting ENABLE_PROMPT_CACHING_1H=1 keeps the one-hour lifetime while you are on usage credits.

A scheduled task fires on its interval while the session sits idle, sending your full context each time, and a message from another of your sessions arrives as a new turn when this one goes quiet, doing the same. Each active agent teammate keeps consuming tokens until it exits. And /compact reads the conversation it summarizes, so a compaction on a big session is one of the largest single requests you can send, and /clear is free.

That last pair is the practical one. Compaction is worth its cost when you need continuity, and it is pure waste when you were only moving to a different problem.

What to do the first time you hit a limit

1. 01 Read the message. Which word appears in it, session, weekly, or Opus, decides everything that follows, and the reset time is printed on the same line.
2. 02 If it says Opus, run /model, pick another model, and carry on. That is the only one of the three a model switch fixes.
3. 03 If it says session or weekly, the reset time is the answer. Nothing you run locally shortens it.
4. 04 Run /usage before the next long stretch, so you learn where you stand while you can still act on it.
5. 05 Run /clear when you switch to unrelated work. Every previous message is resent on every turn, so a session that has wandered through three problems pays for all three on each new one.
6. 06 Point at files by path instead of pasting them. Anything you paste sits in context in full for the rest of the session.
7. 07 Ask for a plan before a big change. A plan costs a few hundred tokens; a wrong 400-line change costs that many times over, twice, plus the turns spent explaining what went wrong.

The last three are Anthropic’s own advice, from article 14552983, and each works for the same reason: it shrinks what gets re-sent on every future turn rather than what you send once.

Keeping CLAUDE.md short belongs on that list for the same reason. Anthropic’s memory documentation describes the mechanism: the file is read once when a session opens and handed to the model as a message ahead of your first one, not as part of the system prompt, so it sits in context for every turn after that. The same page sets the size target under 200 lines and says to add an entry the second time Claude makes the same mistake, and the cost documentation suggests moving detailed workflow instructions out into skills, which load on demand instead of at session start. What belongs in that file is a separate decision from how long it is: standing rules the agent has already got wrong, not everything true about the project.

The two routes that cost money are worth stating plainly rather than burying. /usage-credits lets you keep working past the plan allowance, and what it does depends on who you are. A Pro or Max subscriber gets Settings then Usage on claude.ai, where credits can be switched on or off and the balance and monthly spend limit are visible. A Team or Enterprise member with billing access gets the organization’s Admin settings then Usage. A member without billing access is asked to confirm, then a request goes to the admins. The command is unavailable with API key authentication. The other route is a different plan, which changes the ceiling without changing how the windows work.

Common questions about Claude Code usage limits

Why do Claude Code’s limits feel lower than they used to?

Two effects run in opposite directions. Anthropic doubled the five-hour rate limits and removed the peak-hours reduction on 2026-05-06, so the ceiling moved up. Meanwhile long sessions, agent teammates, and scheduled tasks all spend the allowance without new typing, so the same day of work drains more of it than it used to.

Neither effect is measurable from outside, because no base figure was published to compare against. Your own account is measurable, in the /usage breakdown, where behavior flags mark anything taking 10% or more of recent usage.

Is a Claude Code rate limit the same thing as an API rate limit?

No. They are different systems with the same name. Your subscription window is an allowance measured over hours and a week. An API rate limit is an organization-level throughput ceiling measured per minute, and it applies to a Console or cloud-provider account rather than to a Pro or Max plan.

Anthropic’s API rate limits documentation puts it in one line: “Rate limits set the maximum number of API requests an organization can make over a defined period of time.” That page describes requests per minute, input tokens per minute, and output tokens per minute, set by a usage tier, returning a 429 with a retry-after header. None of that vocabulary applies to the message your terminal printed if you are signed in with a subscription.

Do Claude Code and the Claude app share the same usage limit?

Yes. Claude Code, the browser, and the desktop app all draw on one allowance, so a long research session in the chat product can shorten your coding afternoon without you opening a terminal at all.

Anthropic’s article on using Claude Code with a Pro or Max plan, dated 2026-06-11, states it directly: “Both Pro and Max plans offer usage limits that are shared across Claude and Claude Code, meaning all activity in both tools counts against the same usage limits.” The same page adds that IDE usage counts against those shared limits too. ChatGPT plans are built the same way, and how the two products compare on price and on what runs out sets the two windows side by side.

Is there a Claude Code usage limits dashboard?

Four surfaces exist and none is a dashboard in the reporting sense. /usage inside Claude Code, the rate_limits fields in a custom status line, the usage ring next to the model picker in the Desktop app, and Settings then Usage on claude.ai each show current consumption as bars or percentages.

Coverage separates them. The cost documentation is explicit that /usage figures are approximate and read from local session history on that machine, so Settings then Usage on the web is the only one that counts everything.

Does switching models get me working again?

A model switch rescues one of the three messages. /model restores access after the Opus limit, because that ceiling applies to Opus requests alone. It does nothing for the session limit or the weekly limit, both of which are shared across every model.

The errors reference states that outright. If the message does not say Opus, the reset time it prints is your only real information.

Are the limits different on Pro, Max, Team, and Enterprise plans?

Plan differences are structural in two ways only. Team and seat-based Enterprise members draw on a per-seat allowance sized by their seat tier, Standard or Premium, and a usage-based Enterprise organization has no included per-seat allowance at all and is billed on consumption.

Everything else is the same shape: a rolling five-hour window and a weekly window, shared with the chat product. What differs between Pro, Max 5x, and Max 20x is an allowance size Anthropic does not publish, which is why every plan-by-plan table online is an estimate.

Is the free plan enough to use Claude Code?

The free plan does not include Claude Code, so there is no window to run out. Anthropic’s setup documentation, linked in the plan table above, states that Claude Code requires a Pro, Max, Team, Enterprise, or Console account and that the free Claude.ai plan does not include Claude Code access.

The one route needing no subscription is an Anthropic API key or a cloud provider account. It has no subscription plan window, but requests can still stop at rate limits, exhausted credits, or a configured spend limit.

Can I raise the limit, or buy past it?

Buying past it is supported and raising it is not. /usage-credits turns on metered usage beyond the plan allowance for a Pro or Max subscriber, sends a request to an admin for a Team or Enterprise member without billing access, and is unavailable on API key authentication.

A higher plan is the other route, and it lifts the ceiling rather than removing it. None of the Anthropic pages checked here on 2026-08-16 documents a way to reset a window early or stretch one, so anything promising otherwise is either switching you to metered billing or routing your requests somewhere that is not Anthropic.

Does the plan price tell me what the limit is?

The price tells you what you pay, not what you get. Anthropic documents which windows a plan runs on and which surfaces share them, and publishes no figure for what those windows hold, so two people paying the same amount can do very different quantities of work before either one stops.

That keeps the cost question and the limits question genuinely separate. What each Claude Code plan costs, including per-token billing on an API key has its own answer with its own check dates, and none of it will tell you where your window ends.