How Claude's Usage Limits Really Work in 2026

Claude's limits are not one number. They are rolling caps that interact - a 5-hour session window, a weekly cap covering every model, and on Max a separate weekly Fable meter - and any one of them can stop you mid-task. Anthropic does not publish exact message counts, the caps reset on rolling clocks rather than at midnight, and your Claude Code and claude.ai usage share the same pool. Here is the mechanics, in plain English.

Update, 2026-09-19: Max plans no longer run a separate weekly limit scoped to Opus. Opus and Sonnet are both current, fully available models, and both draw from the single weekly limit that covers every model. The second weekly meter on Max is scoped to Fable instead: Max includes Fable for up to 50% of the weekly limit, after which Fable keeps running on pay-as-you-go usage credits. Pro has one weekly limit across all models and bills Fable from usage credits from the first message. Checked against a live Claude Max 5x account that used Fable, Opus and Sonnet in the same week: Settings > Usage shows current session, this week, and Fable this week, with no Opus meter.

Update, 2026-09-18: Anthropic's Fable models (Fable 5, and Fable 5.1 since 2026-09-01) sit above Opus. On Max plans they are included for up to 50% of your weekly limit at no extra cost, and they drain that limit faster than Opus or Sonnet; past the 50% mark Fable keeps running on pay-as-you-go usage credits. On Pro, Fable is not part of the plan limits at all: every Fable message is billed from usage credits. The Free plan has no Fable access. Everything below about the 5-hour and weekly windows still applies.

The three limits that matter

Anthropic confirms three distinct limit dimensions on every paid Claude plan: a five-hour session window, a weekly usage cap that applies across all models, and a separate weekly cap on Opus. Anthropic's own usage best-practices page spells out the structure but does not publish the numbers - your specific cap depends on plan tier, message length, attached files, tool use, and which model you pick.

Here is the shape of each one, the question it answers, and what trips it.

If you hit a limit, it is one of these three. The error message you see ("usage limit reached") does not always tell you which - and that matters because the next reset is very different in each case. The 5-hour limit can clear in minutes; the weekly limits can leave you waiting days.

The 5-hour session window, explained

The first cap is the one most users notice first. Open a fresh conversation on Monday at 9:00 AM, work through a heavy coding session, and you might hit the 5-hour limit by lunch. It is the cap that makes long, dense days feel suddenly hostile.

The crucial detail: it is rolling, not clock-based. There is no "your 5-hour window starts at midnight." It starts when you send your first message of a new window. If you sent that message at 9:00 AM, your window closes at 2:00 PM. If you sent it at 10:30 AM, it closes at 3:30 PM. After it closes, you get a fresh budget.

This explains a piece of Claude folklore that confuses new users: the "rolling" trap. Someone runs out at 11:00 AM, assumes the limit will reset at 5:00 PM (5 hours later from when they ran out), and is surprised to still be locked out at 5:00 PM. The reset clock counts forward from your first message of the window, not from when you hit the cap.

What burns through a 5-hour window the fastest:

• Long conversations. Every message you send includes the full prior context. A 30-message thread costs more per turn than a 3-message thread, even if the latest message is identical.
• Opus by default. Opus consumes the budget faster than Sonnet, which consumes it faster than Haiku. The cheaper-model path of "Sonnet for exploration, Opus for the hard parts" is the single largest budget-saver heavy users use.
• Tool use and code execution. Each tool call counts. Heavy Claude Code sessions with file reads, web fetches, and bash executions can burn the budget in fewer messages than you would expect from a pure chat session.
• Attached files. Every file you attach is tokenised into the message. A 500 KB PDF can be the equivalent of a long message on its own.

The weekly limits: the new pain in 2026

Until early 2026, hitting Claude's limits usually meant tripping a 5-hour session cap. Heavy users on Pro learned to pace their bursts. Then Anthropic added a weekly all-models cap on top, and later a separate weekly Fable meter on Max. (Max also ran a separate, tighter weekly Opus cap for a time; that meter no longer appears, and Opus draws from the all-models weekly cap like every other model.) These are the limits that catch heavy daily users by surprise - especially Pro and Max 5x subscribers who used to skate through.

Both weekly caps run on a 7-day rolling window per account. They are not "Monday to Sunday." They are "the last 168 hours starting from your first message." That means there is no calendar day on which "everyone's limit resets" - your reset depends on the rhythm of your own usage.

What changed in 2026, and why Reddit and Hacker News lit up:

• The all-models cap is shared across every model. Switching to Haiku or Sonnet to "save tokens" still spends against the weekly bucket. The only way to extend a weekly window is to send fewer messages, not cheaper ones.
• The Fable meter is separate and tighter. Max 5x and Max 20x subscribers expect a big budget; Fable is included only up to 50% of the weekly limit before it starts billing usage credits. The fix is to default to Sonnet and reserve the heavier models for the parts where they actually win.
• Claude Code shares the pool. Anthropic states this directly: "Your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit." If you used to think of Claude Code as a side channel, the 2026 limits will surprise you.

"People are hitting usage limits in Claude Code way faster than expected. We're actively investigating - it's the top priority for the team." Anthropic, March 2026 (via The Register)

Anthropic acknowledged the wave of limit complaints publicly on 2026-03-31 and identified three causes: peak-hours quota reductions affecting roughly 7% of users, the end of a promotion that had doubled limits, and a prompt-caching bug that could inflate token costs 10 to 20 times. The promotion expired on 2026-03-28; the other two are ongoing realities of the 2026 limit landscape. Source: The Register, 31 Mar 2026.

Why your Claude Code session burns faster than claude.ai

Two reasons, both about density.

First, Claude Code does more per message. A single "implement this feature" prompt can trigger dozens of file reads, multiple bash executions, web fetches, and back-and-forth tool calls. Each one counts against your budget. From the user's point of view, you sent one message. From the limit's point of view, you spent a small session.

Second, Claude Code conversations carry more context. The CLI loads the entire `CLAUDE.md` file (often hundreds of lines), the relevant project files, sometimes a full directory tree. Every subsequent message in that session includes all of that as input. By the time you are 10 turns deep, each new turn is paying for the cumulative context of all 10.

This is the mechanism behind the "I said hello and used 13%" reports from 2026: the user typed one short message, but the CLI loaded the entire project context to handle it, and Anthropic's prompt-caching bug at the time (acknowledged in March 2026) made the cache lifetime collapse, so every message paid the full context cost over and over.

What you can do about it

• Start new threads aggressively. Long Claude Code conversations are expensive conversations. A new conversation drops the cumulative context. If you have just finished a unit of work, start a fresh one.
• Use /compact when a session feels heavy. Claude Code's compact command rolls the conversation history into a summary, reducing per-turn input cost from there on.
• Pick Sonnet for exploration, Opus for the hard parts. Default to the cheaper model and switch up only when you actually need it. Most coding sessions don't.
• Keep `CLAUDE.md` lean. Every line gets loaded into every message. Trim aggressively.

Peak hours and the March 2026 squeeze

Peak-hours throttling is the dimension that does not appear on any plan page. Anthropic reduces effective quotas for a portion of users during the busiest hours - confirmed in their public statement as affecting around 7% of users at any given time. There is no documented schedule for "what counts as peak," and no setting to opt out.

For most users, peak-hours throttling is invisible. For the 7% who get caught in it, it shows up as a session that burns through 5-hour or weekly budget faster than the same workload did the day before. The fix is not a single setting; it is awareness. If your usage feels suddenly meaner, check whether you are running heavy sessions in US business hours and whether shifting to off-peak (evenings UTC, mornings UTC) gives you back some headroom.

This is the dimension that most often surprises power users. The 5-hour and weekly caps are documented; peak-hours throttling is not. You can do everything right - lean prompts, Sonnet by default, short threads - and still find your budget evaporating on a particular Wednesday afternoon because Anthropic's data centres are saturated. A live meter is the only way to catch this in the act. Without one, you tend to blame your own behaviour for what is really a server-side dial moving silently.

How to find out where you actually stand

Anthropic surfaces your current usage in one place: sign in to claude.ai, open Settings > Usage, and you will see your consumption against your plan limits, plus the next reset timestamp if you have been throttled.

It works, but it has the same friction every web settings page does: you have to remember to check it, which means you usually check it after you have run out. The whole point of a meter is to see the needle drop while you can still do something about it.

Three ways to monitor it without opening a browser tab:

1. The Claude.ai settings page on a refresh. Free, official, but requires you to remember.
2. The /usage command in Claude Code. Shows your current pool in the terminal. Useful while you are mid-session in the CLI, but only the CLI.
3. A menu-bar meter. tokn.watch reads your existing claude.ai session cookie, polls the same endpoint the settings page uses, and shows your 5-hour and weekly remaining percentages plus reset countdowns right in your Mac menu bar. Free, no servers, runs across both Claude and ChatGPT if you use both.

What to do when you hit a limit

If you have already hit the 5-hour, the fastest unstuck path is usually to switch models - drop to Sonnet or Haiku - and finish low-stakes work while the window rolls forward. If you have hit a weekly cap, you are looking at hours-to-days, not minutes-to-hours, and the action set is different.

The honest answer: there is no "unlock the limit early" button on the consumer plans. Anthropic does not offer pay-as-you-go top-ups on Pro or Max for chat. The options are:

• Wait until the reset. Annoying but free.
• Switch to a different account or provider. Many heavy users keep a Claude account and a ChatGPT account on standby for exactly this reason.
• Upgrade. If you are tripping limits weekly on Pro, the next plan up is the cleanest fix. Pro → Max 5x is a 5x weekly budget bump; Max 20x is the heaviest consumer plan.
• Use Claude API directly. For the technically inclined, the API has its own (much higher, pay-per-token) limits. Different product, but it shares no quota with your Pro plan.

Frequently asked questions