Claude Code Usage Limits Explained: 5-Hour Windows, Weekly Caps, and How to Stop Hitting Them (2026)

I run Claude Code every day, and the single most common question I see — in the community, and from people who ask me how I work — isn't about prompting or agents. It's "why did I hit my limit already?" The usage system is genuinely confusing: there are two timers running at once, it counts tokens instead of prompts, and the "extra usage" safety net has a way of quietly turning into a bill.

So here's the whole system, laid out. Mechanics are grounded in Anthropic's own docs; the "how people actually cope" parts come from what heavy users report (linked where it matters). If you're weighing plans, this pairs with my Claude Code Max vs ChatGPT Pro comparison — that one is about which plan; this one is about how the limits work once you're on one.

The Two Timers

Every paid plan runs two limits at the same time, and you can hit either one:

1. A rolling 5-hour window. This is the one you bump into during a heavy session. Critically, it doesn't run on a fixed clock — the counter starts on your first prompt. Send your first message at 9 AM and the window resets at 2 PM, regardless of how much you sent in between.
2. A weekly cap sitting on top of it. Max plans actually have two weekly limits: one across all models, and a separate one for a specific model tier. Weekly limits reset at a fixed day and time assigned to your account — not a rolling window, a fixed schedule.

The 5-hour usage feeds into the weekly total, so a few intense sessions early in the week can leave you rationing by Friday. The reset clock shown in the app is the number to trust — not any "you get N prompts" estimate you read online.

It Counts Tokens, Not Prompts

This is the part that trips people up most. A "message" doesn't have a fixed cost. Claude counts tokens — input, accumulated conversation context, attachments, tool calls, and output length all go into the meter.

The practical consequence: late in a long session, even typing continue can be expensive, because the model has to re-read the entire accumulated context to answer. This is also why the same type of small fix can cost a few hundred tokens one time and several thousand another — the cost tracks context size, not how big your request looks.

Your Usage Is Shared Everywhere

Claude Code doesn't get its own private allowance. Usage is shared across Claude Code, the Claude.ai chat app, and Cowork — all activity counts against the same pool. Burn through it writing code in the terminal, and you'll find the chat app rate-limited too. Worth knowing if you treat the chat as a separate free resource; it isn't.

What Happens When You Hit the Wall

By default, Claude Code hard-stops when you hit the window — even mid-task. When it does, nothing is deleted and nothing is charged. You have three options:

• Wait it out. The 5-hour window resets on its own. Check exactly when with the /status command.
• Drop to a lighter model. /model switches to a cheaper tier that draws down your allowance more slowly — enough to keep moving on simpler work.
• Enable extra usage. The /extra-usage command lets Claude Code keep going past your plan limit, billing the overflow at standard API rates on top of your subscription.

The Extra-Usage Trap

That third option is the one to understand before you turn it on, not after. Extra usage is real pay-as-you-go billing at API rates — separate from your flat subscription fee. You prepay a balance, and you should set a monthly spend cap when you enable it (there's an "Adjust limit" control, plus optional auto-reload and a $2,000 daily redemption ceiling).

The failure mode shows up in the wild: one user ran a long i18n task, hit the limit, had extra credits enabled — and reported the whole credit balance drained while the task still didn't finish. Whether or not that was working as intended, the lesson is the same: a token-metered overflow with no tight cap can spend real money fast on exactly the kind of long, context-heavy task that made you hit the limit in the first place.

How Heavy Users Avoid the Wall

The people who run Claude Code all day mostly stopped fighting the limits and started designing around them. The patterns that recur:

• Spend the expensive model on thinking, not typing. Planning and code review burn far less than implementation. Use the top model to plan and review; route the bulk implementation to a cheaper tier via /model.
• Watch /status, not vibes. The reset clock and remaining allowance are right there. Checking mid-session beats discovering the wall mid-task.
• Run a second tool for overflow. A very common setup is Claude for planning plus a second coding agent (Codex is the usual pick) for implementation and cross-review, so neither subscription's limit is the sole bottleneck. Users keep both in sync by symlinking CLAUDE.md to AGENTS.md so the same project instructions drive both tools — switching between them becomes a toggle, not a migration.

That last pattern is the same conclusion my Max vs ChatGPT Pro comparison landed on from the other direction: two model families that check each other beat maxing out one.

The One Part That Keeps Changing

Here's the honest caveat: the specific numbers move constantly. As I write this in mid-July 2026, the community is actively debating a change to weekly limits and which models are included in subscriptions — promotional boosts ending, a top model's subscription availability in question, dates that people can't even agree on (one widely-discussed thread argues over the 13th versus the 14th). I'm deliberately not repeating any of those figures as fact, because by the time you read this they'll likely be wrong.

What doesn't change is the structure above: two timers, token-based counting, shared pool, hard stop with an extra-usage escape hatch. Learn the structure, use /status for the live numbers, and check the docs before renewing on the strength of any specific limit you read in a Reddit thread — including the ones I linked.

FAQ

Why did I hit my limit so fast?

Almost always because you're deep in a long, context-heavy session. Limits count tokens, not prompts, and every message re-reads the accumulated context — so a long conversation drains the meter far faster than the same work split across fresh sessions.

Does Claude Code have separate limits from the chat app?

No. Usage is shared across Claude Code, Claude.ai, and Cowork — one pool. Heavy terminal use will rate-limit your chat app too.

What's the difference between the 5-hour and weekly limits?

The 5-hour window is a rolling timer that starts on your first prompt and resets 5 hours later — it's the one you hit during intense sessions. The weekly cap is a fixed-schedule limit sitting on top; your 5-hour usage feeds into it. You can hit either.

Is enabling extra usage safe?

It works, but it's real pay-as-you-go billing at API rates, not part of your subscription. Set a monthly spend cap before turning it on — an uncapped overflow on a long agentic task can spend a lot before it finishes.

Bottom Line

The limits aren't as arbitrary as they feel — they're just two timers metering tokens out of one shared pool, with a pay-as-you-go trapdoor underneath. Once you can see that structure, the fixes are obvious: keep sessions short, spend the expensive model on planning, watch /status, and cap your extra usage. The exact numbers will keep changing; the shape won't.