Claude's paid plans run on two limits: a session limit that resets every five hours and a weekly limit across all models. Neither counts messages. They count work, and every message makes Claude read the whole conversation again, along with any files, project knowledge and tool results in it. Most people who run out are paying for four things: chats that run too long, more context than the job needs, more model than the job needs, and rounds of rework. Fix those and the same plan goes a lot further.
How do Claude's usage limits work?
There are two clocks. The session limit runs on a five-hour window and resets on its own. The weekly limit covers every model and resets at a fixed time each week that Anthropic assigns to your account. On Max there's a third: a separate weekly allowance for Fable, Anthropic's top model. The Free plan only has the five-hour limit.
Both clocks draw on one pool. Chats in the browser, the desktop app and Claude Code all count towards the same limits, so a long coding session in the afternoon leaves less for writing in the evening.
Anthropic doesn't publish a message count, because there isn't one. Its own guide lists what uses more: longer messages, bigger attachments, longer conversations, tools such as web search and Research, the model you pick, the effort level, and building artifacts. For scale, it says a Pro user having short conversations on a lighter model gets around 45 messages every five hours, often more. Max gives five or twenty times Pro per session.
To see where you stand, open Settings, then Usage. You'll see how much of the current session you've used and when it resets, and when the weekly limit resets. When you hit a limit you can wait, upgrade (a new plan's limits apply straight away) or switch on usage credits, which keep you working at pay-as-you-go rates instead of stopping.
Why do you run out faster than you expect?
Because the fortieth message in a chat costs far more than the first. Claude doesn't remember a conversation the way you do. Each time you press send, it reads everything above your message again and then answers. A quick question at the bottom of a long thread is a quick question plus the whole thread.
We call this the re-read tax. Here's what it looks like for the same thirty questions asked two ways.
Same answers. About 60% less reading.
One long chat makes Claude read about 279,000 tokens to answer thirty questions, and the last answer alone needs 18,000. Split the same work into three chats of ten, with a short handover note carried into the second and third, and the total drops to about 107,000. That's roughly 60% less reading for the same answers.
Anthropic says as much in plainer words. When a chat gets long enough that Claude starts summarising earlier messages to make room, that chat uses more of your limit, and its advice is to start a new conversation.
You didn't send too many messages. You sent one very long one, thirty times.
Where do your tokens actually go?
We sort it into four things. We call them Intersect's Four Ls, and nearly every "I hit my limit by lunchtime" story is one of them.
Intersect's Four Ls
How long the chat has run.
Claude re-reads the whole chat before every reply, so each message costs more than the last.
One task per chat. Move on with a handover note.
Everything in the chat that isn't your question.
Files, Project knowledge, connector results and switched-on tools all get read too.
Projects for repeat files. Send the part. Tools off until needed.
How hard Claude works on each reply.
Bigger models, higher effort and extended thinking use more per answer.
Match model, effort and thinking to the job.
The rounds it takes to get it right.
Every "no, shorter" is another full trip, re-read included.
Say it all up front. Batch questions. Edit, don't correct.
How do you cut the length?
One task, one chat. When the task changes, start a new one. A chat that began as "summarise this report" and drifted into a pricing debate and then a client email is carrying the report into every email draft.
When you do need to carry work forward, don't carry the whole chat. Ask Claude for a handover note, check it, and paste it into a fresh chat. This is the prompt we use:
We're moving to a new chat. Write a handover note I can paste into it. Include: the goal, what we've decided and why, the latest version of the work (or where it lives), anything still open, and the next step. Leave out anything we tried and dropped. Keep it under 300 words.
On paid plans Claude can also search your past chats and use memory, so you can say "carry on from yesterday's pricing chat" rather than pasting it back in.
How do you lighten the load?
Load is everything in the chat that isn't your question. Four habits take most of it out.
- Put repeat material in a Project. Files you add to a Claude Project are cached, so when you use them again only the new or uncached part counts against your limit. Your brand guide, price list or style rules belong there, not pasted into every chat.
- Send the part, not the pile. If the question is about clause 14, send clause 14. Don't upload the same file twice in one chat either; Claude already has it.
- Switch off tools you aren't using. Anthropic calls tools and connectors token-intensive. Web search, Research and connectors you've left on can add weight to every message. Turn them on per chat from the tools menu when the task needs them.
- Ask for totals before rows. Connected to your CRM, analytics or a spreadsheet? Ask for the monthly totals first and drill into detail only where something looks off. Ten thousand rows pulled into a chat are ten thousand rows Claude reads again on every reply.
Keep Project instructions short too. They ride along with every message in that Project.
How do you match the lift to the job?
The biggest model on maximum effort with extended thinking switched on is the most expensive way to rewrite an email. Anthropic's advice is to turn extended thinking off when you don't need it and to pick a lower effort level for routine work, because higher effort uses more tokens. Here's how we'd set Claude for common jobs:
Match the lift to the job
| Job | Model | Effort | Extended thinking | Tools |
|---|---|---|---|---|
| Rewrite an email, tidy notes, summarise a page | Haiku or Sonnet | Low | Off | Off |
| Draft a proposal or report from your notes | Sonnet | Medium | Off | Off |
| Analyse a contract, model or messy data export | Opus | High | On | Only the connector it needs |
| Check a current fact with a source | Sonnet | Low | Off | Web search |
| A cited report across many sources | Opus | High | On | Research |
How do you cut the loops?
Every "no, shorter" or "that's not what I meant" is a full round trip, re-read tax included. Most loops come from the first message.
- Say it all up front. Who it's for, how long, what it's for, what good looks like and anything Claude shouldn't do. One clear message beats three corrections.
- Batch the questions. Five related questions in one message cost less than five messages, and Claude can see how they connect.
- Edit, don't correct. If Claude misread you, edit your original message instead of adding a correction underneath. Claude answers the edited version on a new branch, and the wrong answer stops being part of what it re-reads.
- Read before you send. It's on Anthropic's own list, and it's the cheapest tip there.
Does the time of day matter?
It has. In March 2026 Anthropic said it was making five-hour sessions on Free, Pro and Max run down faster during weekday peak hours, 5am to 11am Pacific time, with weekly limits unchanged. From the UAE that's the late afternoon and evening. In May it removed the peak-hour reduction for Claude Code on Pro and Max and doubled Claude Code's five-hour limits. If your heavy work is in chat rather than Claude Code, the mornings in the Gulf are the better time to do it.
How do we run it at Intersect?
This blog is written with Claude, and it runs on the Four Ls. Our playbook, design kit and notes sit in a Claude Project, so every post starts from them without anyone pasting them in. Each post gets its own chat: research, draft, a designed preview and rounds of notes, then the draft goes into our website's CMS. When a chat gets long because a post needed a lot of notes, we finish the post there and start the next one fresh. The weekly limit stays a number we glance at, not one we plan around.
Where does this break?
Some work needs the long chat. Reviewing a long contract clause by clause, or rewriting a document against comments from earlier in the conversation, depends on everything above it. Splitting that work saves tokens and loses the thread. Keep it in one chat, put the document in a Project, and accept the cost.
Handover notes drop things. Claude writes a good summary, but it decides what mattered. Read the note before you start the new chat, and add back anything you'd miss.
A smaller model isn't always cheaper. If Haiku gets a hard question wrong and you ask three more times, every retry re-reads the chat, and you can end up paying more than one go with Opus would have cost. Loops cost more than lift.
And sometimes it's the plan. If you've fixed all four and still hit the weekly limit most weeks, you're a heavy user, and that's what Max, Team and usage credits are for.
Six changes, ten minutes
- Open Settings, then Usage, and note when your weekly limit resets.
- Start a fresh chat for your next task. Bring anything you need over with the handover prompt.
- Move the files you paste most often into a Project.
- Switch off web search and connectors in chats that don't need them.
- Drop the model and effort for routine jobs, and turn extended thinking off.
- Next time Claude misreads you, edit your message instead of correcting it.
Want your team getting more from the same plan?
We set teams up on Claude properly: Projects with your house rules and documents, the right model and settings for each kind of job, and an hour with your people on habits like these. If you're already hitting limits every week, we'll tell you whether it's the habits or the plan.
FAQs
Does starting a new chat reset my limit?
No. It resets what Claude has to re-read, not the meter. Your session and weekly limits reset on their own clocks, which you can see in Settings, then Usage.
Do Claude Code and the Claude app share a limit?
Yes. The browser, desktop app and Claude Code all draw on the same session and weekly limits.
What happens when I hit the limit?
Claude tells you when it resets. You can wait, upgrade, or turn on usage credits in Settings, then Usage, so work carries on at pay-as-you-go rates up to a monthly spend limit you set.
Is Max worth it?
If you've fixed the Four Ls and still run out most weeks, probably. Max gives five or twenty times Pro's usage per session. If it's only the odd week, usage credits cost less.
We use ChatGPT, Gemini or Copilot as well. Does this apply?
The Four Ls apply to all of them, but each has different limits and settings. We've written the same guide for each one.
Usage limits, assistant by assistant
Related reading
- How to Prompt ChatGPT and Claude: An AI 101 for Your Team
- Set Up a Shared AI Workspace for Your Team in an Afternoon
- 10 things Claude Design can do that most teams miss
Sources
- Claude Help Center: Understanding usage and length limits, accessed 6 November 2026
- Claude Help Center: Usage limit best practices, accessed 6 November 2026
- Claude Help Center: About Claude's Pro plan usage, accessed 6 November 2026
- Claude Help Center: About Claude's Max plan usage, accessed 6 November 2026
- Claude Help Center: Troubleshoot Claude error messages, accessed 6 November 2026
- The Register: Anthropic tweaks timed usage limits to discourage Claude demand during peak hours, 26 March 2026
- PCWorld: Anthropic confirms it's been adjusting Claude usage limits, 27 March 2026
- Techsy: Claude usage limits in 2026, every 2x change explained, updated 26 June 2026