If you searched for this because Claude cut you off mid-task and told you to come back in a few hours, the honest answer is the one almost nobody puts at the top of the page: Anthropic does not publish a number for the Claude Pro plan. No message count. No hour count. No token quota. Not for the five-hour window, and not for the weekly one.
We checked ten official Anthropic pages for this article, and every figure below is quoted from one of them. Where Anthropic states a rule, we quote it. Where Anthropic states nothing, we write Not published and name the pages we looked at. That sounds like a small distinction, but it is the whole difference between a page you can act on and a page that invents a plausible-looking number and leaves you planning your week around it.
The short answer
Claude Pro costs $20 per month billed monthly, or $17 per month billed annually. In exchange you get, in Anthropic’s own words on the Pro plan page, “At least five times the usage per session compared to our free service.”
Five times what, exactly? Five times an unpublished number. Anthropic anchors the entire consumer plan ladder to a base quantity it has never disclosed. Pro is 5x Free. Max 5x is 5x Pro. Max 20x is 20x Pro. Every rung is a multiplier on a figure that does not appear anywhere in the documentation.
This is not an oversight, and it is not really unreasonable either. Usage on Claude is not measured in messages — it is measured in tokens, and the number of tokens a single message costs varies enormously depending on how long your conversation already is, whether you attached a 40-page PDF, whether Claude ran a web search, and which model you picked. A published message count would be wrong for almost everyone. But it does mean that any page telling you Pro gives you “45 messages every five hours” is telling you something Anthropic has not said.
The three limits Pro users actually run into
Anthropic draws a distinction that most write-ups collapse: usage limits and length limits are two different systems. You can hit either one, and hitting them feels completely different.
1. The five-hour session limit
The Pro plan page is unambiguous about the cadence: “Your session-based usage limit will reset every five hours.”
This is a rolling window, not a fixed schedule tied to the clock. It starts when you start. If you burn through your session allowance at 10am, you are waiting until roughly 3pm, not until midnight.
2. The weekly limit
Sitting above the session window is a second ceiling. The Pro plan page: “Pro plans also have a weekly usage limit that applies across all models.”
The weekly limit behaves differently from the session limit in one important way — it is not rolling. Anthropic states that “Your reset day and time stay the same regardless of when you start using Claude.” The Max plan page puts it the same way: “Weekly limits reset at a fixed time each week that is assigned to your account.”
Two consequences follow that are worth internalising. First, your weekly reset day is assigned to you, not chosen by you, and it is not the same day for every account. Second, the practical failure mode on Pro is not usually the five-hour limit — it is running your weekly ceiling down to nothing by Wednesday and then discovering that the five-hour window resetting no longer helps you, because the weekly ceiling is what is actually blocking you.
How big is the weekly limit? Not published.
3. Length limits are a separate system
The third wall is the context window, and it has nothing to do with your subscription tier’s quota. From Anthropic’s page on how usage and length limits work: “The context window size depends on which model you’re using. On paid plans, the newest models support up to a 1M token context window, while others support 500K or 200K tokens.”
Hitting a length limit means one conversation has grown too large to continue. Hitting a usage limit means you have spent your allowance for the period. Starting a new conversation fixes the first and does nothing for the second.
Pro has one weekly limit. Max has two.
This is the structural difference between the tiers that we have not seen stated anywhere else, and it matters more than the headline multiplier.
The Pro plan page describes a single weekly ceiling: a weekly usage limit “that applies across all models.”
The Max plan page describes something different: “Max plans also have two weekly usage limits: one that applies across all models and another for Sonnet models only.”
So Max is not simply “Pro, multiplied.” It is a different limit architecture, with a model-specific ceiling layered on top of the global one. If you are comparing the plans on price per unit of usage, the second ceiling is a real constraint that the 5x and 20x numbers do not capture. Upgrading does not just raise your ceiling; it changes the shape of the ceiling. We break down what the 5x and 20x multipliers do and do not mean, and how they compare with OpenAI’s equivalents, in our piece on Claude Max vs ChatGPT Pro usage limits.
Claude Code, Cowork and the web app all draw from the same Pro allowance
If you use Claude Pro for both chat and terminal work, this is probably the single most useful fact on this page. The support article on using Claude Code with a Pro or Max plan states that the plans offer “usage limits that are shared across Claude and Claude Code, meaning all activity in both tools counts against the same usage limits.”
Anthropic’s own Claude Code cost documentation extends the same principle to seat-based plans, noting that the per-seat allowance “is shared with Claude chat and Cowork.”
There is no separate Claude Code quota to discover, and no way to keep an agent session from eating the allowance you wanted for chat. A long autonomous coding run and an afternoon of conversation are drawing from one pool. If you have been surprised by how fast Pro runs out, and you use Claude Code, this is almost certainly why. For how that shared pool compares with what OpenAI includes for Codex on a ChatGPT plan, see our Codex vs Claude Code comparison.
What actually burns through a Pro session
Anthropic’s usage limit best practices page, last updated 2 June 2026, lists the factors that determine consumption. The six named drivers are message length, file attachment size, current conversation length, tool usage such as Research and web search, model choice, and artifact creation and usage.
The page on how usage and length limits work adds one more that is easy to overlook: “Your usage is affected by several factors, including the length and complexity of your conversations, the features you use, which Claude model you’re chatting with, and the effort level you’ve selected.” Effort level is a setting, not a consequence of what you asked for — which makes it one of the few consumption levers you can pull deliberately.
Note the third item on the list: current conversation length. Because every turn re-sends the conversation so far, a long thread costs more per message than a short one, even if your messages are identical. A forty-turn conversation is not forty times as expensive as a one-turn conversation — it is worse than that, because the cost of each turn grows as the thread does. Splitting a long working session into fresh conversations at natural boundaries is the cheapest habit change available to a Pro user.
The counterweight is Projects, and Anthropic is explicit about the mechanism: “When you upload documents to a project, they’re cached for future use”, and “Every time you reference that content, only new/uncached portions count against your limits.” If you keep re-pasting the same specification, style guide or codebase excerpt into new conversations, you are paying for it every time. Put it in a Project’s knowledge section and you pay for it approximately once.
The Fable 5 carve-out: not included in Pro
This is the exception that breaks the “Pro gets a share of everything” mental model, and it is recent enough that most pages ranking for this query predate it.
Anthropic’s page on Claude Fable 5 states that the model “is available on all paid plans (Pro, Max, Team, and Enterprise).” Available, however, does not mean included. For Pro subscribers and standard Team seats, the page says plainly that “Fable 5 isn’t included in your plan’s usage limits” — access runs on pay-as-you-go usage credits instead.
Max subscribers and premium Team and Enterprise seats get a different deal: “you can use up to 50% of your weekly usage limits on Fable 5 at no extra cost.” And the model is expensive to run in either case, because it “draws from your plan’s regular weekly usage limits and uses them faster than other Claude models.”
The practical reading for a Pro subscriber: your $20 does not buy you Fable 5. If that is the model you want, the relevant comparison is not Pro versus Max on quota size, it is Pro-plus-credits versus Max-with-a-50%-allowance.
What happens when you hit the limit
Two paid escape hatches exist, they are priced differently, and the difference is worth real money.
Usage credits
Anthropic’s page on managing usage credits, last updated 18 May 2026, confirms availability on “Pro, Max 5x, and Max 20x” plans. Credits are billed as “consumption-based pricing at standard API rates”, which means you stop being a subscriber at the margin and start being an API customer.
There is a ceiling: “There is a daily redemption limit of $2000.” Auto-reload is available, described as the ability to “automatically add funds when your balance drops below a threshold you set.” Whether credits expire is Not published — the page does not address it.
Usage bundles
Bundles are the same thing bought in advance at a discount, and unlike almost everything else in this article, they come with real numbers. Anthropic’s usage bundles page, also updated 18 May 2026, confirms “Usage bundles are available on Pro, Max, and Team plans.”
| Bundle value | You pay | Discount |
|---|---|---|
| $50 | $45 | 10% |
| $250 | $200 | 20% |
| $1,000 | $700 | 30% |
Source: support.claude.com, Buy usage bundles, updated 18 May 2026.
Ordering matters: “Bundle credits only apply after you’ve exceeded your plan’s limits.” A bundle does not raise your ceiling, it pays for what happens past it. And there is a monthly cap on the discount: “Individual Pro and Max plan subscribers can purchase up to $2000 worth of discounted bundles per month”, with Team owners allowed $3,000 per month.
If you routinely spend past your Pro allowance, the arithmetic is stark. Paying as you go at standard rates costs full price. The same spend routed through the $1,000 bundle costs $700. On a $2,000 monthly overage that is a $600 difference for the same work, and the only thing you changed was buying in advance.
The cache lifetime trap
Here is a detail buried in Anthropic’s Claude Code cost documentation that we have not seen surfaced anywhere in the search results for this query. Prompt caching — the mechanism that makes repeated context cheap — does not behave the same way once you are on credits. The cache lifetime is an hour on a subscription, and “drops to five minutes once you’re drawing on usage credits.”
Read that again in terms of your bill. The moment you exhaust your Pro allowance and start paying per token, the thing that was keeping your long conversations affordable gets twelve times shorter. Work that fit comfortably inside a one-hour cache window now misses the cache repeatedly. Your effective cost per message does not just go from zero to something — it goes to something noticeably higher than the same message would have cost under subscription caching.
The implication is that credits are worse value than the headline API rate suggests, and that the correct comparison for a heavy user is between buying credits and upgrading the subscription, not between credits and nothing.
How to see your own usage instead of guessing
Since Anthropic will not tell you the size of your allowance, the next best thing is measuring your own consumption against it. Claude Code ships a command for exactly this.
Running /usage shows your recent consumption, and per Anthropic’s documentation you can press “d” or “w” to switch between the last 24 hours and the last 7 days. The 7-day view is the one to watch, because it maps onto the weekly limit that is most likely to be what actually stops you.
Two caveats from the same documentation. The figures are “approximate and computed from local session history on this machine” — so if you work across a laptop and a desktop, neither machine sees the whole picture. And on paid plans the view goes further than a single total: it attributes recent usage to “skills, subagents, plugins, and individual MCP servers”, and flags behaviours such as long context or cache misses when they account for at least 10% of recent usage.
That attribution is the most useful diagnostic available to a Pro user. If a single MCP server or a chatty plugin is quietly consuming a third of your weekly allowance, this is the only place it shows up. Two related commands round it out: /usage-credits opens billing settings on Pro and Max, and /cost reports running session spend for API-key users.
Is Pro enough?
Without published quotas, nobody can answer this from the documentation, including us. What the documentation does support is a decision procedure.
Pro is likely sufficient if your work is conversational, your threads are short, you are not running Claude Code for hours at a stretch, and you do not need Fable 5. At $17 to $20 a month it is inexpensive relative to a single hour of most people’s time, and the five-hour reset means an occasional limit is an interruption rather than a wall.
Pro is likely insufficient if you use Claude Code daily, because of the shared pool; if you work in very long threads with large attachments; or if you want Fable 5, which Pro does not include at any usage level. In those cases the honest options are Max at $100 per month and up, Pro plus bundles, or a different tool entirely — we survey the field in Claude Code alternatives and compare per-token pricing across vendors in our AI coding assistant pricing breakdown. The crossover is easy to compute: Pro at $20 plus $80 a month of bundle-discounted credits is $100, which is exactly what Max 5x costs. Past that point you are paying Max money without the Max allowance, and you are doing it with the shorter cache window.
The one thing we would not advise is planning around a message count you read somewhere. It is not a number Anthropic has published, so it is not a number anyone can hold them to.
Every page we checked
For transparency, and so you can verify any of this yourself, these are the official sources behind this article and what each one contributes.
| Source | What it establishes | Last updated |
|---|---|---|
| support.claude.com — What is the Pro plan? | Five-hour reset; 5x Free; one weekly limit; fixed reset day | 10 June 2026 |
| support.claude.com — How do usage and length limits work? | Usage vs length distinction; context window sizes; effort level | Not dated precisely |
| support.claude.com — Usage limit best practices | Six consumption drivers; Project caching behaviour | 2 June 2026 |
| support.claude.com — What is the Max plan? | 5x and 20x multipliers; two weekly limits on Max | Not dated precisely |
| support.claude.com — Use Claude Code with your Pro or Max plan | Shared pool across Claude and Claude Code | 11 June 2026 |
| support.claude.com — Manage usage credits for paid Claude plans | Credits on Pro and Max; API rates; $2,000 daily cap; auto-reload | 18 May 2026 |
| support.claude.com — Buy usage bundles | Bundle prices and discounts; $2,000 monthly cap; ordering rule | 18 May 2026 |
| support.claude.com — Claude Fable 5 on your plan | Fable 5 excluded from Pro limits; 50% allowance on Max | Not dated precisely |
| claude.com/pricing | $17 annual / $20 monthly for Pro; Max from $100 | Not dated |
| code.claude.com — Claude Code costs | /usage command; cache lifetime drop on credits | Not dated |
Three of those pages carry no visible last-updated date at all, which is worth knowing when you are relying on them for a purchasing decision.
What is still missing
Four things a Pro subscriber reasonably needs, and none of them are published today.
The size of the weekly limit in any unit — tokens, messages, or a percentage readout inside the product for chat users as well as Claude Code users. The base quantity that the 5x and 20x multipliers multiply, without which the multipliers are decorative. Whether purchased usage credits expire. And a per-model consumption rate, so that “model choice affects usage” becomes actionable rather than directional.
We will update this page when any of those appear. Until then, the most accurate thing anyone can tell you about Claude Pro’s usage limits is the shape of the system, the list of things that make you hit it sooner, and an honest Not published everywhere a number should be.
Moving up from Pro? See our comparison of Claude Max 5x vs 20x to pick the right tier, and check whether Claude Code is free before you upgrade.
Already hit the wall? When Claude usage limits reset covers the five-hour session window and the fixed weekly reset time assigned to your account, quoted from the official help pages.
Running out of room inside one conversation rather than across the week is a length limit, not a usage limit. What fills that window, and how to get it back, is covered in the Claude Code context window.
Anthropic publishes multipliers and no absolute figures. OpenAI does the opposite: see Codex usage limits by tier for the per-model local-message allowances, and whether Claude Code is free for what the entry tiers actually cost.
Update, 22 August 2026: the temporary 50% weekly uplift, and what happens on 1 September
Everything above describes Anthropic’s published limits. Since May 2026 the weekly limits actually in force for Claude Code have been running above those published figures, and that gap was scheduled to close at the end of August 2026. It did not close then; it was extended twice more, and as of 14 September 2026 Anthropic’s own support documentation still contains no record that it ever existed. See the 14 September update at the foot of this page before you rely on any uplift figure.
On 19 August 2026, Anthropic’s developer channel posted: “We’re extending the 50% increase to weekly Claude Code limits through August 31. We hope to make this a permanent change to our plans, but strong demand for our models means that capacity may be tight over the coming weeks. We’ll keep you posted as things develop.”
The uplift applies to Pro, Max, Team and seat-based Enterprise plans. The details that matter:
- It affects the weekly window only. The rolling five-hour window is unchanged, so if you are hitting session limits rather than weekly limits, this uplift was never helping you and its expiry will not hurt you.
- It is not reflected in the published documentation. Anthropic’s support articles still describe the standard weekly bands. There is no page where you can look up the boosted number, which is why the practical answer to “what is my weekly limit” remains the one this article gives: check
/usagein Claude Code, because the documented figure and the enforced figure are currently different. - It has been extended repeatedly rather than made permanent. Anthropic has stated it hopes to make the change permanent, while citing capacity pressure as the reason it has not committed. That is a genuine caveat, not a formality.
If you have been sizing your plan during the uplift period, the conservative planning assumption is the published weekly limit, not the one you have been experiencing. That is the number that is guaranteed to you. Anyone whose workflow only fits inside the boosted allowance should know that it is a promotion Anthropic has explicitly declined to guarantee.
We will update this page when Anthropic states what happens after 31 August. In the meantime, if the weekly ceiling is the binding constraint on your work, the comparison worth running is against OpenAI’s published Codex allowances, which are documented per model and per tier — see Codex usage limits, every published number by tier and our side-by-side on Claude Max vs Codex usage limits. If you have not committed to a paid plan yet, is Claude Code free covers what the free tier does and does not include.
Update, 31 August 2026: today is the last day of the uplift window, and there is still no follow-up announcement
The 50% weekly uplift described in the section above was extended “through August 31” — that is today. We checked Anthropic’s website, news posts and support documentation on 31 August 2026 and found no follow-up announcement stating either that the uplift has been made permanent or that it is ending. The 19 August developer-channel post quoted above remains the most recent statement we can source. We are not going to guess which way it goes; if you depend on the weekly headroom, the honest planning assumption is that it may revert without further notice, and the observable test is your own usage meter next week rather than any published figure.
Worth noting because the two get confused: a different Anthropic temporary arrangement was resolved today, and resolved favourably. The $2/$10 per million token rate for Claude Sonnet 5, announced at launch as introductory pricing through 31 August 2026, is now stated in Anthropic’s pricing documentation as the standard price, and the scheduled increase to $3/$15 on 1 September 2026 will not occur. That is an API pricing change and has no bearing on subscription usage limits, which is what this page is about — a Pro or Max plan’s weekly window is not denominated in dollars per token. We cover the pricing side on our AI coding assistant pricing page and in is Claude Code free.
Update, 14 September 2026: the 50% weekly uplift has expired, and no Anthropic support page ever documented it
The temporary uplift described in the two updates above was last extended to 13 September 2026. That date has now passed. On 14 September 2026 we re-read Anthropic’s support article on usage limits for Claude and Claude Code, word for word, to see how the expiry was being communicated. The answer is that it was not communicated, because the uplift was never in that article in the first place.
The article, which carries an “Updated over 3 weeks ago” stamp, describes only that there are “usage limits that are shared across Claude and Claude Code” and lists the four things you can do once you hit them: upgrade your plan, turn on usage credits, move to the Console and buy API credits, or wait for the reset. That is the whole of it. There is no numeric ceiling anywhere on the page. There is no weekly figure. There is no mention of a 50% increase, no mention of an extension, and no mention of 13 September.
The full lifecycle happened outside the documentation
It is worth being precise about what that means, because it is unusual. A capacity change went through its entire life — announcement, first extension, second extension, expiry — without ever appearing in the vendor’s own product documentation. The 19 August wording quoted above came from a developer channel post, not a support page. The extension past 31 August reached us through an in-product screenshot, which is a user-visible surface but not a published source. And the expiry has no announcement at all: the uplift did not end with a notice, it ended by the date passing.
So we are changing how this page treats the number. From today, the 50% uplift is marked here as non-official and expired. We are leaving the 19 August quotation in place, because it was accurately transcribed on the day and deleting it would falsify our own record, but it should now be read as history rather than as a figure to plan against. If you are sizing a week of Claude Code work, do not budget on the uplift.
The transferable rule: for capacity, the documentation is not the source of truth
Yesterday we recorded a closely related finding on a different vendor: Z.ai published an end date for a discount, and the discount was still live when that date arrived, so the published end date turned out to be a statement of intent rather than a billing event. Today’s finding is the sharper sibling of that one. There, the published date was wrong. Here, there is no published date, and there never was.
Put the two together and the rule is the same in both directions. For subscription capacity, the vendor’s documentation is not the source of truth. Your own usage meter is. You cannot open a support article and learn whether you currently have an uplift, because the uplift was never there to read. What you can do is watch where your own week actually runs out, and treat any figure that reached you through a blog post, a developer channel, or a screenshot as provisional until your account behaves the way it says.
This is also why we keep saying that per-token list prices and subscription ceilings are two different kinds of number with two different reliabilities. List prices are published, versioned and checkable — we re-read Anthropic’s pricing page again on 14 September and every line was unchanged for the sixth consecutive check. Capacity is none of those things. The gap between how well those two are documented is, in our experience, the single most common reason people’s expectations about a coding subscription turn out wrong.
The same expiry affects the comparison on our Codex usage limits page, which quoted the identical 19 August wording, and it is relevant to the reset mechanics we describe in how Claude usage limits reset and to the tier comparison in Claude Max against Codex on usage limits. All three have been checked against the same 14 September reading.
Checked 14 September 2026 against Anthropic’s official support article on Claude and Claude Code usage limits. Where we say a figure is absent, we mean it did not appear in that article on that date.
Update, 20 September 2026: one coding subscription now publishes its weekly allowance in tokens, and that makes it checkable in a way message counts never are
Every question that brings readers to this page is some version of “how much can I actually run for what I am paying?” On 20 September 2026 we recorded a competitor answering that question in a unit that can be audited, and the contrast is worth a section — not because one vendor is cheaper, but because of what unit each one chose.
Z.ai’s Coding Plan documentation publishes an “Estimated Token Allowance” table: for each plan tier and each assumed cache hit rate, a range of millions of tokens per week. The Lite tier at a 95% cache hit rate is stated as 48–97M tokens per week on GLM-5.3; the Pro tier as 290–580M; the Max tier as 676–1,352M. We have written up the full table, including where the ranges come from and what the vendor’s accompanying savings claim does and does not support, in our GLM Coding Plan vs Claude Code comparison.
A wide number in the right unit beats a precise number in the wrong one
The instinct is to treat a 2x-wide range as worse information than a crisp figure. For this particular question, it is the opposite, and the reason is falsifiability.
A weekly allowance denominated in tokens can be checked against something you already have. Your API logs, your editor’s usage panel and your own context windows are all counted in tokens. If a vendor says a plan yields roughly 290 to 580 million tokens a week and your own instrumentation says you burned 400 million and were not throttled, the claim survived a test. If you were throttled at 120 million, it failed one. Either way you learned something.
An allowance denominated in messages, prompts or sessions cannot be tested that way, and not because the vendor is being evasive. It is arithmetic: a message has no fixed size. A limit of “N messages per window” is compatible with almost any token volume, because one message can be a two-word follow-up and the next can drag a 200,000-token repository into context. Two users on the identical plan, both stopping at exactly the same message count, can differ by orders of magnitude in what they actually consumed — and neither of them has been told anything false.
This is the structural reason the “how many messages do I get” question that fills forums never resolves. The unit does not carry the information the asker needs. The answers are not inconsistent because people are guessing; they are inconsistent because the quantity being reported genuinely is not the same quantity from one user to the next.
The practical consequence: what you can and cannot forecast
If your allowance is stated in tokens, you can forecast a month ahead from last month’s logs, and you can tell in advance whether a new workload fits. If it is stated in interactions, you cannot forecast at all — you can only discover the ceiling by hitting it, which is why the common experience of a subscription coding tool is a limit that seems to move: it is not moving, your average request size is.
That gives a usable rule for anyone comparing subscription coding plans, and it is independent of which vendor you prefer: convert every plan you are considering into tokens per week before comparing anything. If a vendor does not publish that number, the honest entry in your spreadsheet is “unknown”, not an estimate you derived from a message count. An unknown you can go and measure over one billing period. An estimate built on an unstated average request size will look authoritative and be wrong by a factor you cannot bound.
Two warnings before anyone divides one plan by another
First, and this one is decisive for cross-vendor comparison: tokens are not a standard unit. Anthropic states that Claude 4.7 and later use a tokenizer that produces approximately 30% more tokens for the same text. So the same source file counted under two vendors’ tokenizers is two different numbers, and a weekly token allowance from one vendor cannot be divided by another vendor’s to produce a ratio. It tells you what you can run on that vendor. It does not tell you how that compares.
Second, published allowance ranges may have conditions attached that have nothing to do with your code. In Z.ai’s case the width of the range is entirely a function of a peak/off-peak schedule defined in UTC+8, which means the number a reader should use depends on their timezone rather than their workload. That is covered in detail on the comparison page linked above; the general lesson here is that a range’s width usually has a stated cause, and finding it is normally more useful than averaging the two ends.
Related reading on this site: the full weekly-token table and its conditions in GLM Coding Plan vs Claude Code, the per-token API rates behind every subscription in AI coding assistant pricing, and the tier-by-tier comparison in Claude Max vs Codex usage limits.
Correction, 24 September 2026: the Anthropic sentence quoted above is no longer on the vendor’s page
This page cites Anthropic’s explicit statement that a previously scheduled increase to $3 and $15 per million for Sonnet would not go ahead. That statement was on the official pricing documentation when we wrote it and was still there on 23 September 2026. As of 24 September 2026 the sentence has been removed from the page, and the rate table has been restructured.
The prices themselves are unchanged: Sonnet 5 is $2 input and $10 output per million, cached input $0.20. Budget against those figures, because they are what is published and what you will be billed.
We are leaving the original passage in place with this correction attached rather than editing it away, and we are not predicting a price rise — a note that has served its purpose is a routine thing to remove in a page redesign, and we do not publish guesses about vendor intent. We explain the change, and why archiving vendor prose with a read-date matters more than archiving vendor tables, in our AI coding assistant pricing comparison.