The Claude Opus usage limit

Opus has its own weekly ceiling, separate from the weekly cap that covers all models. That is why you can be locked out of Opus on a Wednesday while everything else still works perfectly, and why switching models gets you moving again in seconds rather than making you wait for a reset. Opus usage also draws down the general weekly cap at the same time, so it is two meters running at once, not one.

Why Opus has a cap of its own

A Claude subscription meters you on a rolling five-hour session window and on weekly caps. The weekly side is really two ceilings: one that applies across every model you use, and a separate one that applies to Opus only. Your usage screen shows them as distinct rows with their own progress bars and their own reset times.

The reason is cost, and it is not mysterious. Anthropic's current lineup runs from Claude Haiku 4.5, the fastest model, through Claude Sonnet 5, described as the best combination of speed and intelligence, to Claude Opus 5 for complex agentic coding and enterprise work, with Claude Fable 5.1 above that for demanding reasoning and long-horizon agentic work. On the metered API path, the closest public proxy for relative cost, Opus 5 lists at $5 per million input tokens and $25 per million output tokens against Sonnet 5 at $2 and $10, and Haiku 4.5 at $1 and $5.

A single flat weekly cap would mean a subscriber who defaults to the most expensive model consumes several times what a Sonnet user does for the same nominal allowance. A dedicated Opus ceiling keeps the expensive model available on every paid plan without letting it swallow the plan. The practical effect: Opus is a resource to spend deliberately, not a default to leave switched on.

Opus draws down two meters at once

This is the part people miss. Time spent on Opus does not come out of a ring-fenced Opus budget that leaves everything else untouched. It counts against both the Opus weekly ceiling and the weekly cap across all models, and it counts against your current five-hour session window as well.

So there are three ways an Opus-heavy week can end badly, and they need different responses:

• Opus cap full, general cap fine. The common case. Switch models and carry on immediately; there is nothing to wait for.
• General weekly cap full. Switching models does not help, because that ceiling covers all of them. You are waiting for your weekly reset. This is the outcome that heavy Opus use makes arrive earlier than it otherwise would.
• Session window full. Also model-independent, but measured in hours rather than days.

Which is the honest argument for not leaving Opus as your default: it is not only that you will run out of Opus, it is that you will reach your general weekly ceiling sooner than a Sonnet-first week would have taken you there. Every Claude surface feeds the same meters too, so a browser conversation on Opus and a terminal session on Opus compound.

One related wrinkle if you use Fable models: they have no separate cap. On Max they count toward your ordinary weekly limits, with up to half your weekly allowance usable on them at no extra cost; on Pro they run on pay-as-you-go usage credits instead.

What "Opus limit reached" looks like, and the instant fix

In Claude Code the message is literally "You've hit your Opus limit", and it sits alongside three siblings that mean quite different things: "You've hit your session limit", "You've hit your weekly limit", and "You've hit your Sonnet limit". The rule that separates them is worth memorizing:

• Model-specific limits (Opus, Sonnet) are escapable. Switch to a model outside that family with /model, or the model picker in the app, and you keep working immediately.
• Session and weekly limits apply across all models, so switching does nothing and you wait for the reset.

That means the correct first move on seeing an Opus message is not to check your plan, open a support ticket, or start a new conversation. It is to press /model, pick Sonnet, and keep going. On a long unattended run, recent versions of Claude Code will also hold the session open and continue the interrupted task automatically once a limit resets, and switching models during that wait causes it to re-check and resume early. More on the full set of messages in Claude usage limit reached.

When Opus earns it, and when Sonnet is the better default

Anthropic's own guidance for Claude Code is direct: Sonnet handles most coding tasks well and costs less than Opus, so reserve Opus for complex architectural decisions and multi-step reasoning, and consider Haiku for simple subagent tasks. That maps cleanly onto real work.

Worth Opus:

• Architectural decisions where a wrong call is expensive to unwind later.
• Genuinely hard debugging, where the failure crosses several systems and the cause is not where the symptom is.
• Long-horizon planning for a change that will touch many files, especially in plan mode before any code is written.
• Work where you will act on the answer without checking it closely, so the quality of the first attempt is the whole game.

Sonnet by default:

• Routine implementation against a plan you have already agreed.
• Refactors, test writing, and mechanical edits across known files.
• Exploration and codebase reading, where volume matters more than depth.
• Anything you were going to review carefully anyway.

A pattern that works well: plan on Opus, execute on Sonnet. Planning is a small share of the tokens and a large share of the value, so the expensive model is spent where it changes the outcome, and the long tail of implementation runs on the cheaper meter. Doing it the other way around, which is what leaving Opus as the permanent default amounts to, spends your scarcest ceiling on the work that needed it least.

Seeing the Opus meter and planning around the reset

Open Settings > Usage in the Claude web or desktop app, or run /usage in Claude Code. The Opus ceiling appears as its own row, with a progress bar, a percentage from 0 to 100, and its own next reset time in your local timezone. /usage in Claude Code returns immediately without interrupting a response, so you can check it mid-task. Walkthrough: how to check your Claude usage.

Once you know your reset, heavy Opus work becomes schedulable rather than a gamble:

• Front-load the cycle. Weekly caps reset at a fixed time each week assigned to your account. Put the architecture and the hard debugging in the first days after your reset, while the Opus row is full. See when Claude limits reset.
• Treat the last day of the cycle as Sonnet time. Mechanical work does not care which model does it.
• Never start a long unattended Opus run against a nearly full bar. The run will stop mid-task, and the wait is days rather than hours if it is the general weekly cap you meet.
• Set the default deliberately. /model switches mid-session and /config sets the default. Choosing once beats deciding on every task.
• Watch the warning. Claude flags a window as it gets close, which is a pacing signal rather than an alarm: what the approaching weekly limit warning means.

The Opus bar you never fill

Here is the other thing that separate row on the usage screen tells you. Both weekly ceilings are ceilings, not balances. At your reset they refill to the same level whether you used all of the Opus allowance, some of it, or none of it, and nothing accumulates. Plenty of people who worry about the Opus cap discover, once they look, that they finish most weeks with both bars well short of full.

Tokens for Good turns that recurring surplus into something useful. Your Claude claims a queued nonprofit, researches its real-world impact against a fixed methodology with citations, and submits a structured report. Every organization is researched twice by independent contributors, then validated, consolidated, scored deterministically, and human-reviewed before it reaches the public directory. It runs on the subscription you already pay for with no separate API cost, it is scoped to fit inside a single run, and it can run on a schedule so it fills the quiet part of your cycle rather than competing with your own work. "Tokens" here means AI model tokens, not crypto: no coin, no wallet, no blockchain. See how the research works, read what Tokens for Good is, or open the docs.

Frequently asked questions