Last updated: July 2026
Claude Code on a Claude subscription uses rolling rate limits, not a fixed credit pool. Usage is capped on a rolling weekly basis, with shorter session limits on top. Hit a ceiling and you wait for the window to reset or upgrade. Heavy context and long autonomous runs burn the limit faster, so tighter sessions go further.
If Claude Code suddenly tells you to slow down, you have not bought a broken product — you have hit a rate limit. The subscription does not sell you a pile of credits; it gives you access that is metered on rolling windows. Understanding how that meter works is the difference between constantly bumping the ceiling and comfortably getting a full week of work out of your plan.
The confusion starts with the mental model. People assume a subscription buys a monthly bucket of tokens they spend down like a prepaid card. Claude Code does not work that way. Access through a Claude subscription is governed by rolling rate limits: usage is capped over a rolling weekly window, with tighter session-length limits layered on top.
When you reach a ceiling, Claude Code does not silently overcharge you — it pauses and tells you to wait for the window to reset, or to upgrade for more headroom. Because the window rolls, the limit refills gradually as older usage ages out rather than resetting on a fixed calendar date. That is why a heavy morning can leave you throttled in the afternoon even though the "week" is not over.
One honest caveat up front: the exact numbers move. Anthropic tunes tiers and thresholds over time, so this guide focuses on how the system behaves, not on a fixed quota you can memorize. For the current plan tiers and what each one includes, see our Claude Code pricing guide. Below are the four things that actually determine how far your limit goes.
The single biggest source of surprise is expecting a credit pool and getting a rolling window instead. With a pool, you can front-load a huge job and coast; with a rolling limit, a burst of intense usage counts against you for the rest of the window. There is both a shorter session limit — protecting against a single marathon run — and a longer rolling weekly limit that governs sustained work.
Practically, this means pacing matters. Two hours of tightly scoped work spread across a few sessions stresses the limit far less than the same token volume dumped into one sprawling autonomous run. The window is always healing, so steady, deliberate use is rewarded and bursty, unscoped use is penalized.
How much room you get is set by your plan. Pro at $20/month has the tightest ceiling — it is designed for lighter, intermittent coding rather than all-day autonomous agent work. Higher subscription tiers raise the rolling limits substantially, giving heavy users far more headroom before they ever see a throttle.
Beyond the subscription tiers, the pay-as-you-go API is a different model entirely: instead of a rolling ceiling you pay per token with no weekly cap, which suits workloads that would constantly bump against subscription limits. If you are consistently hitting the wall on Pro, the fix is usually to move up a tier or shift the heaviest work to the API — not to fight the limit. Our breakdown of AI coding agent costs compares the subscription and API paths so you can see which fits your usage.
Rate limits are measured in tokens, and tokens accumulate in ways that are easy to miss. Every turn of an agent session re-sends the accumulated context: the files it has read, the tool output it collected, and the full conversation history. A 4,000-token file dump from early in a session gets re-transmitted on turn after turn, so a session that "felt short" can consume a surprising slice of your window.
Long autonomous runs compound this. When you let the agent grind on an under-specified task, it reads files it does not need, retries, and self-corrects — and every one of those loops carries the full, growing context. A single unscoped run across twenty files can eat more of your limit than a whole day of tight, targeted edits. This is exactly the dynamic covered in our guide to reducing the Claude Code context window: shrinking what rides along on each turn is the highest-leverage way to stay under the ceiling.
Not all tokens cost the same against your limit. Opus-class models are far heavier than Sonnet, so the same task run on Opus draws down your window much faster. Reaching for the most powerful model on every routine edit — renaming a variable, fixing a typo, writing a small test — is one of the quickest ways to hit a ceiling early in the week.
The efficient move is to match the model to the job: let Sonnet handle the large majority of routine work and reserve Opus for genuinely hard reasoning or architecture. Bloated context on the heaviest model multiplies the two most expensive variables at once, which is the classic recipe for an unexpected throttle. If you want to see how prompt and context size translate into tokens before you send, our token calculator makes the math visible.
None of this means Claude Code is stingy — it means the limits reward discipline. A handful of habits keep you comfortably inside your window most weeks:
/compact and /clear. Use /compact to summarize and shrink a long conversation, and /clear to start fresh when you move to an unrelated task. Both cut the accumulated history that would otherwise re-ride on every turn.If you want to go further on the prompt layer specifically, Terse compresses what you type into Claude Code before it is sent and tracks per-request token cost inline — a small, on-device nudge that trims waste so each rolling window goes further. It is a light touch, not a rewrite: the goal is simply to stop paying for words the model never needed.
Terse compresses the prompts you send into Claude Code and tracks per-request token cost — on-device, zero latency, no API calls. Trim the waste before it eats your rolling window.
Terse for Claude CodeClaude Code on a Claude subscription uses rolling rate limits rather than a fixed credit pool. Usage is capped on a rolling weekly basis, with shorter session limits on top. When you hit a ceiling you wait for that window to reset or upgrade to a higher tier. Heavy context and long autonomous runs consume the limit faster.
Rate limits are driven by tokens, not messages. Long autonomous runs, large files, and accumulated context re-sent every turn burn through your window fast — especially on Opus. A single unscoped agent run over many files can consume more of your limit than a full day of tight, focused edits.
Use Sonnet for routine work and save Opus for hard reasoning, run /compact and /clear to shrink conversation history, scope tasks narrowly, avoid redundant file reads, and batch work into focused sessions. Trimming prompt and context waste stretches how far each window goes.
Claude Code pauses until the rolling window resets, or you can upgrade to a higher tier for more headroom. Pro has the tightest ceiling; higher subscription tiers and the pay-as-you-go API give you more room. Exact numbers change over time, so treat the mechanism, not a fixed quota, as the thing to plan around.