Claude usage limits, explained
If you have ever been cut off after twenty messages one day and run for hours the next, nothing is broken and your account is not throttled. This page explains the metering that produces that result, and the handful of habits that move the wall meaningfully further away.
Short answer
Claude does not count your messages. It counts tokens — roughly, the amount of text moving through the model — over rolling windows measured in hours, with a longer secondary window on top for sustained heavy use. That single design choice explains nearly every confusing report about Claude limits: two people sending the same number of messages can get completely different caps, because one was pasting in documents and the other was typing sentences.
The cap on each plan
Read the window column before the cap column. A cap that resets on a rolling window behaves nothing like one that resets on the 1st of the month, and it is the reason two people on the same plan report completely different experiences.
Free
$0Cap
A modest token budget per window, sized so that a short working session finishes but a document-heavy one does not.
Window
Rolling, measured in hours from your first message of the session
When you hit it
Chat input is disabled and the interface shows the time your allowance returns. Existing conversations stay readable.
Pro
$20/moCap
Several times the free budget per window, plus access to the larger models. Anthropic publishes this as a multiple of Free rather than as an absolute number, precisely because the real figure depends on what you send.
Window
Rolling hourly windows, with an additional longer-horizon cap for sustained heavy use
When you hit it
Input is disabled until the window rolls forward. Hitting the longer-horizon cap is rarer and takes considerably longer to clear.
Max
$100–$200/moCap
Published as a multiple of Pro — the two Max tiers exist specifically to serve people running agentic coding sessions that a Pro budget cannot sustain.
Window
Same rolling mechanic, much larger budget
When you hit it
Same behaviour as Pro. Even Max is finite: a long-running autonomous coding session will find the ceiling eventually.
Published figures read off Claude's own documentation in August 2026. Caps on this category of product change without announcement and are often adjusted per-account under load, so treat the mechanics below as the durable part and Anthropic's help-centre article on usage limits as authoritative over anything on this page.
What actually consumes one unit
This is the part that lets you predict your own cap instead of being surprised by it.
Tokens, not messages
Roughly three-quarters of a word per token, counted on both what you send and what Claude sends back. A one-line question costs almost nothing. A pasted 30-page contract costs more than a hundred ordinary messages, in a single send.
Every turn re-reads the entire conversation
This is the mechanic that surprises people most. In a long thread, each new message carries the whole transcript with it, so cost per message climbs steadily as the conversation grows. Message forty in a thread can cost twenty times what message two did.
Attachments stay in context for the rest of the thread
Uploading a large PDF does not cost you once — it costs you on every subsequent turn in that conversation, because the document is re-read each time. This is the fastest known way to exhaust a window.
Agentic and coding sessions consume quota in the background
Tool-using sessions generate many model calls per instruction you type. One 'refactor this module' can be dozens of internal turns, which is why heavy coding users exhaust budgets that chat users never approach.
Extended thinking costs real output
Reasoning modes produce substantial internal output before the visible answer, and that output is metered. Leaving extended thinking on for trivial questions is quietly expensive.
How the reset really works
- •Windows are rolling and start at your first message — not at midnight, and not at a fixed hour. If you began at 8:40am, that is when your window began.
- •The interface tells you the exact time your allowance returns. That in-product timestamp is authoritative for your account; no published table will be.
- •There is a second, longer-horizon cap layered on top of the hourly one, aimed at sustained heavy usage rather than single sessions. Most chat users never encounter it; people running long agentic coding sessions do.
- •The budget is shared across surfaces on the same account, so heavy CLI or IDE use consumes the same allowance the web app draws on.
- •Model choice matters. Lighter models draw far less from the budget for the same conversation, so dropping down for routine work extends the window substantially.
Getting more out of the plan you already pay for
Try these before upgrading. For most people one or two of them push the cap far enough out that the next tier stops being necessary.
Start a new conversation the moment the subject changes
By far the highest-leverage habit. Long threads are quadratically expensive: you pay for the whole history on every turn. Carrying forward a three-sentence summary into a fresh chat costs a rounding error compared to keeping the thread alive.
Do not leave large documents sitting in an active thread
Extract the sections you actually need, work on those, and start a clean conversation for the next document. An attached PDF is a recurring charge on every turn, not a one-off.
Use a lighter model for routine work
Formatting, summarising, renaming, tidying — none of these need the largest model. Reserving it for genuine reasoning can double or triple how long a window lasts.
Turn extended thinking off for simple questions
Reasoning modes are worth their cost on hard problems and pure waste on lookups. Leave it available, not always on.
Ask for the diff, not the file
In coding work, asking Claude to return only the changed lines rather than reprinting a whole file cuts output tokens dramatically, and output is the more expensive half of the meter.
If the wall only ever arrives during coding sessions, moving that workload to a request-metered coding tool is cheaper than a bigger assistant tier.
Upgrade, or move?
Pay for the higher tier when…
- ✓You are running agentic coding sessions daily — that workload is exactly what the Max tiers were built for, and no prompting habit will bring it inside a Pro budget.
- ✓You hit the wall mid-task more than twice a week on work you bill for. The lost context on a forced stop typically costs more than the tier difference.
- ✓You genuinely need the largest model throughout, so dropping to a lighter one for routine turns is not an option.
Look elsewhere when…
- →Your usage is bursty rather than sustained — a second $20 subscription elsewhere buys more usable overflow capacity than one larger tier does.
- →What you actually need is sourced, current answers off the live web, which is a different product with much looser metering at the same price.
- →The cap only ever lands during coding. A request-metered coding tool absorbs that workload on its own budget and leaves your chat allowance intact.
Where the quota is looser
- ChatGPT — Meters per-model message allowances rather than tokens, which is more forgiving of long conversations and less forgiving of frequent short re-prompts. Its limits page is linked below.
- Cursor — Purpose-built for coding and metered in requests, so agent runs stop competing with the chat budget you also want for writing and research.
- Perplexity — For research-shaped questions with citations, a much cheaper way to buy volume than upgrading an assistant tier.
Common questions
Why do I hit Claude's limit faster on some days than others?
Because the meter counts tokens rather than messages. A day spent pasting documents into one long thread consumes vastly more than a day of short standalone questions, even if the message count is identical. Thread length is the other half of it — every turn re-reads the whole conversation, so a long thread gets progressively more expensive.
How long is the Claude reset window?
The primary window is measured in hours and is rolling — it starts at your first message rather than at a fixed clock time. There is also a longer-horizon cap layered on top for sustained heavy usage. The interface shows the exact time your allowance returns, and that timestamp is the only figure guaranteed accurate for your account.
Does starting a new chat actually reset anything?
It does not restore your budget, but it stops the bleeding, which matters more. A fresh conversation drops the accumulated transcript you were re-paying for on every turn, so each subsequent message costs a fraction of what it did in the old thread. It is the single most effective habit change available.
Do Claude Code and the web app share the same limit?
Yes — the budget belongs to your account, not to a surface. An afternoon of agentic coding will visibly shorten what is left for the web app, which catches out people who think of them as separate products.
Is Claude Max worth it over Pro?
It is if you run agentic coding sessions most days; that is the workload the tier exists for and the one that reliably exhausts Pro. If your usage is bursty rather than sustained, a second subscription to a different assistant usually buys more practical headroom per dollar than a single larger tier does.
Related reading
People hitting Claude's limits ask AI what to move to. Is that your tool?
"I keep hitting the Claude cap, what should I use instead?" is one of the highest-intent questions asked of an assistant — the reader is already paying and already frustrated. ChatGPT, Claude and Perplexity answer it from pages like this one. If you build a tool with a more generous quota, check whether they name you.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.