Notes on coding agents and the pages they leave behind, on moving between notes apps, and on the plain mechanics of keeping what you think. Written and published from a Sumibako vault.
Claude usage limits explained: the five-hour session, the weekly cap, and why they run out
How Claude's usage limits work in 2026: the five-hour session, the weekly cap, what eats them fastest, and what to do when you hit one.
Prashant Shaw
Founder, Sumibako
7 min read
Claude does not tell you how many messages you get. It tells you when you have had them, usually mid-sentence, usually on the one afternoon you needed it. The number is not a secret so much as not a number: in Anthropic's words, how much you can do depends on what you do with it, "so there's no fixed message count."
What Anthropic does publish is enough to stop the limit being a surprise: two clocks, one shared pool, a list of what drains it, and four ways out when it runs dry. Everything here is from Anthropic's own help pages and pricing, checked on 22 October 2026. I build a notes app, which comes up once near the end.
⏳ How Claude's limits work: two clocks
Every Claude plan has a session limit that resets every five hours. Paid plans add a weekly limit on top. From the pricing page: "Every plan has usage limits that reset on a rolling five-hour session window, and paid plans add weekly limits on top."
The weekly clock is the one people misread. It does not start when you subscribe or when you first send a message. The Pro plan article says "Weekly limits reset at a fixed time each week that is assigned to your account", and that the day and time "stay the same regardless of when you start using Claude or when your subscription begins". So if your week resets on a Thursday evening, it resets on a Thursday evening, whatever you did on Monday.
Two more things that are true on every plan:
One pool. "Your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit", per
An afternoon in Claude Code is an afternoon less in the chat.
No message count. What you get is a multiple of somebody else's allowance, not a number of your own.
📊 What each plan gets
Plan
Price
Per five-hour session
Weekly
Free
$0
"Free covers everyday questions"
None published
Pro
$20 a month, or $17 a month billed yearly
At least 5x Free
One weekly limit across all models
Max 5x
$100 a month
5x Pro
Weekly limit, plus a separate one for Fable
Max 20x
$200 a month
20x Pro
Weekly limit, plus a separate one for Fable
From the pricing page: "Pro gives you at least 5x more usage per 5-hour session than Free. Max gives you 5x or 20x more usage per 5-hour session than Pro." The Max article gives the prices and confirms the weekly limit "applies across all models". On Pro, Fable is not part of the allowance at all: the pricing table lists it as "Usage credits".
Note what the multiples are of: the session. Anthropic says Max gets five or twenty times Pro's usage "per 5-hour session", and says nothing about the weekly limit growing by the same factor. Both plan pages also keep a door open: "we may limit your usage in other ways, such as weekly and monthly caps or model and feature usage, at our discretion."
If you are comparing, I put ChatGPT's limits on one page the same way. OpenAI publishes more numbers, though rarely the one people ask for.
🔍 How much you have left
Settings, then Usage. On Pro, Max, Team and seat-based Enterprise, Anthropic's best practices page describes "progress bars showing how much of your five-hour session and weekly usage limits you’ve consumed", the time left in the current session, and when the weekly limit resets.
Before you hit a limit, Claude warns you, per its error messages page: "Approaching session limit" or "Approaching weekly limit". At the limit, without usage credits, you see either "5-hour limit reached - resets [time]" or "You've reached your weekly limit. It resets [day] at [time]."
Claude Code is kinder mid-task. If the session limit arrives while it is partway through a response, it "may keep working briefly so it can finish the step in progress", showing "Usage limit reached · wrapping up", according to the Claude Code plan article. That last bit still counts toward your week. Kinder, sometimes: on Pro this happens at most once a week, on Max each time you reach a session limit, and it is still rolling out, so your account may not have it yet.
🍽️ What eats it fastest
Anthropic's best practices page lists what affects usage: message length, file attachment size, current conversation length, tool usage such as Research and web search, model choice, effort level, artifacts, and multi-step tasks like running code, creating files or browsing websites. Three of those deserve a sentence each.
Long conversations. Every message in a long chat is answered with the whole chat in view, so the hundredth message costs more than the first. Automatic context management makes this worse rather than better: "Longer conversations that trigger automatic context management consume more of your usage limit."
Tools and connectors. "Tools and connectors are token-intensive", the limits article says, and suggests turning off the apps a conversation does not need.
Effort and model. "Higher effort uses more tokens." The usage credits page recommends Haiku, or the latest Sonnet, for most tasks, at "lower token costs than Opus models".
🧯 How to make it last
1️⃣ Start a new chat for a new topic. Anthropic's own advice, from its usage credits page: "Begin fresh chats for new topics to minimize the size of your context window." Carry the useful part across with a short handoff note instead of the whole conversation; I wrote a template for that.
2️⃣ Put reference documents in a project. "Content in projects is cached and counts less against your limits when reused." The catch: "Caches expire after a period of inactivity", so the first message after a long break pays full price.
3️⃣ Match the effort to the job. Turn extended thinking off and lower the effort level for routine work, and save the biggest model for the hard part.
4️⃣ Switch off what you are not using. Connectors, web search, Research. Each one you leave on is paid for in tokens whether or not the answer needed it.
5️⃣ Batch your questions. Five related questions in one message cost less than five round trips, which is the least glamorous tip on the list and possibly the most effective.
6️⃣ In Claude Code, check one environment variable. If ANTHROPIC_API_KEY is set, Claude Code uses it "instead of your Claude subscription" (Anthropic), which means API charges instead of your plan's usage. Useful if you meant it. Expensive if you did not.
Wait. The session resets within five hours, the week on its fixed day.
Upgrade. "If you upgrade, your new plan’s limits apply right away."
Turn on usage credits. On Pro and Max, your requests "switch to pay-as-you-go pricing on usage credits and keep working", billed "at standard API pricing rates", until a monthly spend limit you set or until your balance runs out. The usage credits page has the settings. Note that the monthly API credits included with Max "don’t apply to usage credits in Claude, Claude Code, or Claude Cowork."
Use a limit reset, if you have one. Anthropic gives these out "occasionally to eligible plans". A reset puts your session or weekly limit back to full when you press "Reset for free", on the web or in Claude Desktop.
🐢 "It runs out faster than it used to"
You are not imagining the conversation, at least. Since late September, several open reports in the Claude Code repository describe weekly limits draining faster for the same work: one measures it at about 3.6 times after the reset of 25 September, another at roughly 50 percent more, and a Max 5x user reports a jump since mid-October. When I checked on 22 October, none had a reply from Anthropic.
Before you add your voice, rule out the causes Anthropic does document: a conversation that has grown long enough for automatic context management, a higher effort level or a bigger model than before, connectors left on, or a Claude Code session using your plan where it used to use an API key. If none of those changed, the issues above are the place to add numbers, since numbers are what get a report read.
📍 The cheapest message is the one you do not resend
Nearly every tip above is the same tip: give Claude less to re-read. A fresh chat that starts from a one-page brief costs a fraction of a two-hundred-message thread that drags its history along.
That brief has to live somewhere, which is where I come in, so weigh this paragraph accordingly. I keep mine as pages in Sumibako, the notes app I build, which Claude can read and write on every plan, Free included, through a custom connector. A file, a project or a pinned note works too. The point is that the brief outlives the chat, so the next chat can start small.
Claude's limits are a buffet with no menu prices and a bouncer who appears at the dessert table. Now at least you know where the plates are.
Eight eMudhra keys were integers under 256, one of them 4. Why browsers blocked them in hours, why your Docker image did not, and what to check by 27 October.
