How do I manage Claude Code rate limits?
Manage Claude Code rate limits with /usage, /clear, /compact, model and effort choices. Official allowance sizes remain unpublished.
Claude Code subscription sessions share a rolling five-hour allowance and a weekly allowance with Claude chat and Desktop. Parallel terminals and agent teammates draw from that shared pool. You cannot plan around an official message count because absolute allowance sizes are [GAP]. Watch /usage, reset context between unrelated tasks, and reserve expensive models or high effort for work that needs them. The linked Support and Claude Code docs are the source of the numbers below.
/usage as the live source of truth. Use /clear between unrelated tasks, /compact when you need to preserve context, and Sonnet or lower effort for routine parallel work. At the ceiling, wait for the reset, use usage credits if enabled, or use the documented session auto-continue option. Do not invent message counts or hours of Opus.What limits apply
| Auth path | What you hit |
|---|---|
Subscription /login | Rolling five-hour and weekly allowances, plus model-family bars where shown. Claude chat, Claude Code, and Desktop share the pool. |
| API key (Console, Bedrock, Vertex, or Foundry) | Per-token billing and possible organization or provider RPM, ITPM, or OTPM throttles that can return 429. Use /cost. |
| Context or auto-compact warning | Conversation length, not a plan allowance. |
Support documents the shared usage model in Models, usage, and limits in Claude Code. It does not publish an official message count, token budget, or number of “hours of Opus” for Pro, Max, Team, or seat-based Enterprise. That is a deliberate [GAP], so use /usage in the CLI or Settings → Usage rather than a third-party estimate.
What changed in September 2026?
The weekly-limits promotion says Claude Code weekly limits were 50% higher from May 13 through September 13, 2026. That promotion applied to Claude Code only, while the five-hour window and chat or Cowork limits were unchanged. Starting September 14, 2026, weekly Claude Code limits became permanently 25% higher than before the promotion for Pro, Max, Team, and seat-based Enterprise. Hacker News discussion of ~17% versus the promotion peak is community discourse only, not an Anthropic figure.
How parallel work burns the pool
The practical rule is simple: concurrency changes how quickly the shared allowance is consumed, not how much allowance each terminal receives.
| Pattern | Lower-burn move |
|---|---|
| Multiple terminals | Stagger heavy jobs, stop finished sessions, and keep /usage visible. |
| Agent teams | Keep teams small, use Sonnet teammates for routine work, and shut down finished agents. The Claude Code costs guide describes plan-mode agent teams as costing about 7× standard usage. That multiplier is a cost pattern, not a promise about a fixed rate-limit allowance. |
| One long open session | Run /clear between unrelated tasks. Use /compact mid-task when you must keep the current thread. The usage limit best practices guide explains why a growing conversation can use more of the allowance. |
| Opus or high effort | Use /model for model choice, /model opusplan when that fits the task, and lower /effort for routine work. See how do I set maxEffortLevel in Claude Code. |
Large CLAUDE.md files or pasted context | Keep instructions lean and link to files by path. See how should I share AGENTS.md and CLAUDE.md. |
A useful runbook is: start with /usage, split unrelated work into short sessions, assign routine subtasks to Sonnet, then reserve Opus or high effort for the decision that needs it. Agent teams are off by default and require CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1; if you enable them, treat each teammate as another active context and stop it when its work is done.
What to do at a ceiling
| Message or signal | Action |
|---|---|
| Five-hour or weekly limit | Read the reset time in the message, check /usage, and wait for the applicable reset. /usage-credits may be available on eligible accounts. |
| Model-family limit | Try another available model family with /model. This does not restore a shared five-hour or weekly hit. |
| Spend or credits cap | Check Admin Settings → Usage. Extra usage is described in Extra usage for paid Claude plans. |
| Context or auto-compact warning | Use /compact to keep the task, or /clear if the next task is unrelated. |
API 429 | Treat it as an organization or provider throttle. Back off and retry according to that provider's guidance. It is not the subscription plan bar. |
Claude Code 2.1.234 and later document /rate-limit-options and autoContinueAtUsageLimit for session auto-continue. They do not turn a weekly allowance into an unlimited pool. Desktop weekly auto-continue beyond the linked documentation is [UNVERIFIED]. If you need version-specific CLI context, see Claude Code 2.1.273 to 2.1.278.
Switching models can help only when the message identifies a model-family bar. It does not erase shared usage already consumed. Similarly, /clear reduces future context resend, but it cannot refund usage that has already been counted.
Pitfalls worth avoiding
- Inventing a cap. Support does not publish absolute Pro, Max, Team, or seat-based Enterprise allowance sizes here. Keep the [GAP] and follow
/usage. - Promoting the HN number. The ~17% comparison is discourse only. The sourced product claim is +50% during the promotion and +25% permanently versus pre-promotion from September 14.
- Using
/modelas a reset button. It may address a model-family bar, but not a shared session or weekly hit. - Confusing context-full with rate-limited.
/compactand/clearaddress context; they do not change the plan's allowance. - Leaving expensive parallel work running. Finished teammates, idle check-ins, Opus, and high effort can make a long run end sooner. Use the smallest setup that answers the question.
- Guessing when the five-hour window starts. Support documents a rolling five-hour window, but its exact start trigger is [GAP].
FAQ
Do parallel terminals get separate limits?
No. Subscription sessions share the five-hour and weekly bars with Claude chat and Desktop. Parallel terminals and teammates consume that shared pool; they do not create a new allowance for each process.
How do I tell a session limit from a weekly limit?
Read the limit message and check /usage or Settings → Usage. Those surfaces show the relevant bars and reset times. The exact allowance sizes remain unpublished, so do not convert a bar into an invented message count.
Did Claude Code weekly limits drop in September 2026?
The promotion's +50% period ended September 13. From September 14, Support says weekly Claude Code limits are permanently +25% versus the pre-promotion level. The ~17% comparison with the promotion peak comes from community discourse, not Anthropic.
What is the highest-leverage habit for long runs?
Use /clear between unrelated tasks and /compact when you need the current context. This limits repeated context in future turns. It does not refund usage that was already counted.
Sources
Checked September 21, 2026. No hands-on CLI allowance measurement is claimed here; absolute limits remain a documented [GAP].
- Weekly limits promotion, permanent +25% versus pre-promotion from September 14
- Models, usage, and limits in Claude Code, shared limits and usage surfaces
- Usage limit best practices, session management
- Understanding usage and length limits, context versus usage limits
- Extra usage for paid Claude plans, usage credits
- Manage costs effectively, agent-team cost pattern
- Discourse only: Hacker News discussion