Claude Code usage limits, explained

Claude Code does not have its own limits. On a Pro or Max subscription it draws on exactly the same two clocks as claude.ai and Claude Desktop: a rolling five-hour session window and weekly caps, one across all models plus a separate one for Opus. What changes in the terminal is the rate. Agentic work carries far more context per request than chat does, so the same plan reliably feels smaller in Claude Code than it does in a browser tab.

Which limits apply to Claude Code?

On a subscription, three ceilings govern your terminal work, and none of them are specific to Claude Code:

• The rolling five-hour session window, which opens with your first message and runs about five hours from there.
• The weekly cap across all models, which resets at a fixed time each week assigned to your account rather than a shared calendar hour.
• A separate weekly cap for Opus, which you can hit while the general one still has plenty of room.

The single most important consequence: every Claude surface counts against the same pool. Anthropic states this plainly, that usage across claude.ai, Claude Code, and Claude Desktop all counts toward the same usage limit. So a heavy refactor in the terminal and a long research conversation in the browser compound against one ceiling, and the number on your usage screen already includes both. Full mechanics: Claude weekly usage limits, explained.

Claude Code is included on Pro and on both Max tiers; it is not part of the free plan. On Team and Enterprise plans the allowance is a per-seat pool shared with chat and Cowork, sized by seat tier. If you authenticate with an API key instead, none of this applies: you are metered per token, with no windows to run out of.

Why a single Claude Code session can burn through a window

Anthropic's own guidance for administrators is blunt about this: budget more for a coding seat than a chat seat, because each Claude Code turn carries file contents, tool calls, and multi-step reasoning, and one debugging session can consume more than a day of chat. Here is where it actually goes:

• Long context. Claude Code sends your full conversation with every request, and each tool use sends another request carrying that batch of tool results. A one-line question in a session that has been open all day still draws usage proportional to the whole history.
• Cache misses. Prompt caching keeps re-reading that history cheap, but the first message after a break longer than the cache lifetime reprocesses the full context at full rate. That lifetime is an hour on subscription usage and drops to five minutes once you are drawing on usage credits.
• Tool calls and file reads. Every grep, every file read, every MCP call is a round trip with the conversation attached.
• Subagents and agent teams. Each one runs its own context window. Anthropic puts agent teams at roughly seven times the tokens of a standard session when teammates run in plan mode, because every teammate is a separate Claude instance.
• Retries and wrong turns. A run that heads down the wrong path and gets corrected pays for both attempts. Plan mode exists largely to avoid that.
• Background draw. Scheduled tasks fire on their interval even while the session sits idle, each time sending your full context.

None of this is waste; it is what makes agentic work work. But it explains why a plan that feels generous in chat can feel tight after two long terminal sessions. The habits that cut it most: how to reduce Claude Code token usage.

Checking where you stand with /usage

You never have to leave the terminal to find out. /usage returns immediately without interrupting a response, and shows:

• A progress bar and percentage for each window, plus the exact next reset time in your local timezone.
• An attribution breakdown of what has been consuming your allowance recently: skills, subagents, plugins, and individual MCP servers, each as a share of the total, with a day or week toggle.
• Behavior flags for patterns such as long context or cache misses when one accounts for a large share of recent usage, each with a tip.
• A session block at the top with token counts and a dollar figure computed locally at list rates. On a subscription that figure is informational; it matters when you are billed by API key.

/cost is an alias for the same screen. One caveat: the attribution figures come from local session history on the machine you run the command on, so work from another device or from claude.ai is not reflected there. The plan usage bars themselves do cover your whole account. Walkthrough: how to check your Claude usage.

Model choice is the other lever worth pulling before a big run. Sonnet handles most coding tasks well and costs meaningfully less than Opus; Haiku is faster and cheaper again and is a good default for simple subagents. Reserve Opus for the genuinely hard architectural calls, and remember it has its own weekly ceiling: the Claude Opus usage limit.

Usage limits and rate limits are not the same thing

Worth one paragraph because the words get used interchangeably and mean different things. A usage limit is a subscription ceiling: your five-hour window or a weekly cap is spent, and it refills at a scheduled reset. A rate limit is a throughput control on the API path, expressed in requests and tokens per minute, and it clears in seconds rather than hours. In Claude Code you will see the second one as "Server is temporarily limiting requests" or a bare 429, which Claude Code retries automatically and only surfaces once retries are exhausted. Different again is a 529 Overloaded capacity error, which is Anthropic's servers being busy across all users and, in Anthropic's words, does not count against your quota. If you build against the API as well as using the subscription, the Claude rate limit error untangles the whole set.

Planning long or unattended runs

Long agentic jobs are exactly where limits become a real scheduling problem rather than a curiosity, because the run stops mid-task rather than between tasks. A few things make that manageable:

• Front-load the week. Read your weekly reset time once from the usage screen, then put the heaviest multi-file work in the first days of your cycle while both weekly ceilings are full. See when Claude limits reset.
• Let Claude Code wait through a reset. Recent versions keep the session open when a limit stops Claude mid-task, show a line such as Usage limit reached · continuing automatically at 3:45pm · esc to cancel, and pick the task up on their own afterwards. /rate-limit-options controls the behavior.
• Scope the run so it can finish. A job with a clear end state survives an interruption far better than an open-ended one. How long Claude Code can run covers the practical ceilings.
• Put recurring work on a schedule instead of remembering it. How to schedule Claude Code walks through it, and scheduled runs land naturally in the quiet part of your cycle.
• Clear between unrelated tasks. /clear costs nothing and stops you paying for yesterday's conversation on today's question.

If your Claude Code sits idle most of the week

Here is the pattern that surprises people once they start reading /usage regularly: even developers who feel limit-constrained on their two big days a week finish most weeks with the bar well short of full. That capacity is a ceiling, not a balance. Nothing accumulates, nothing carries over, and every reset refills to the same level whether you used it or not.

Tokens for Good is built for exactly that gap. Your Claude Code claims a queued nonprofit, researches its real-world impact against a fixed methodology with citations, and submits a structured report, then stops; it is deliberately scoped to fit inside a single run. Every organization is researched twice by independent contributors, then validated, consolidated, scored deterministically, and human-reviewed before it reaches the public directory. It installs as an MCP server, runs on the subscription you already pay for with no separate API cost, and can run on a schedule so it quietly fills the windows you were never going to use. "Tokens" here means AI model tokens, not crypto: no coin, no wallet, no blockchain. See how the research works, or the docs, or start with what Tokens for Good is.

Frequently asked questions