Claude Code rate limits: the 5-hour window and weekly caps explained

Three limits, not one

When Claude Code stops accepting work, it names which ceiling you hit. The three messages look almost identical and mean very different things.

You've hit your session limit · resets 3:45pm
You've hit your weekly limit · resets Mon 12:00am
You've hit your Opus limit · resets 3:45pm

The rolling window is the part people get wrong

It is not a bucket that empties at a fixed time. It is a trailing measurement of what you consumed over the past five hours. If you burned most of your allowance between 9am and 11am, that consumption ages out gradually through the early afternoon rather than all at once. Sitting and waiting for a "reset moment" is the wrong mental model; capacity is already returning while you wait, and the error prints the exact time it clears.

What actually consumes your allowance

Limits are measured in compute, which tracks tokens processed, not messages sent. Five things dominate, roughly in order of size.

Reading the gauges before you get blocked

The failure mode worth designing around is not the block itself. It is being blocked twenty minutes into a refactor with a half-applied change on disk. Three surfaces tell you where you stand.

1. /usage, on demand

On a Pro, Max, Team, or Enterprise plan, /usage shows plan usage bars plus a breakdown of what is consuming them: usage attributed to skills, subagents, plugins, and individual MCP servers, each as a percentage. It also raises behavior flags for anything accounting for 10% or more of recent usage, such as long context or cache misses. Press d or w to switch between the last 24 hours and the last 7 days.

2. The status line, continuously

A custom status line receives session JSON on stdin, and that JSON carries the live rate-limit state. This is the cheapest way to keep both windows in front of you at all times.

{
  "rate_limits": {
    "five_hour":  { "used_percentage": 23.5, "resets_at": 1738425600 },
    "seven_day":  { "used_percentage": 41.2, "resets_at": 1738857600 }
  },
  "context_window": {
    "used_percentage": 8,
    "remaining_percentage": 92,
    "context_window_size": 200000
  }
}

#!/bin/bash
# save as ~/.claude/statusline.sh, chmod +x, then point settings.json at it
input=$(cat)
h5=$(echo "$input" | jq -r '.rate_limits.five_hour.used_percentage // 0')
d7=$(echo "$input" | jq -r '.rate_limits.seven_day.used_percentage // 0')
ctx=$(echo "$input" | jq -r '.context_window.used_percentage // 0')
printf '5h %.0f%%  |  7d %.0f%%  |  ctx %.0f%%' "$h5" "$d7" "$ctx"

3. Every device, all the time

What changed in 2026

The limits landscape moved several times, which is why advice written in 2025 misleads now.

How to stop hitting them

Match the model to the task

Use the biggest model for architecture, hard debugging, and anything where being wrong is expensive. Use a mid-tier model for implementation against a clear spec. Use the smallest for mechanical work: renames, boilerplate, running and reading tests. For subagents, set model: haiku in the subagent config.

# Inside a session
/model            # switch model, saved as the default for new sessions
/effort low       # low, medium, high, xhigh, max

# Or start pinned
claude --model sonnet --effort medium

Turn thinking down, not off

Extended thinking is on by default and its tokens bill as output tokens. For simple work, lowering the effort level with /effort is the cheap fix. On models with a fixed thinking budget you can also set MAX_THINKING_TOKENS, for example MAX_THINKING_TOKENS=8000; adaptive-reasoning models ignore nonzero budgets, so use effort levels there instead.

Clear rather than compact between unrelated tasks

/compact has to read the conversation it summarises, so compacting a large context is itself a large request. When the next thing you type is unrelated to the last thing, /clear is both cheaper and better: it costs nothing. Rename the session first if you want to find it again.

/rename tax-rounding-fix   # label it before you leave
/clear                     # free, drops everything
/compact Focus on the API surface and test output   # keeps continuity, costs a pass
/resume                    # come back to it later

Keep tool output small

Tell the agent to pipe noisy commands through head, run focused tests rather than whole suites, and grep with tight patterns. A PreToolUse hook that rewrites test commands to show only failures turns tens of thousands of tokens into hundreds, deterministically, without relying on the model to remember.

Move long instructions out of CLAUDE.md

CLAUDE.md loads into context at session start, every session. Detailed workflow instructions sitting there are paid for even when you are doing something unrelated. Move them into skills, which load only when invoked, and keep CLAUDE.md under about 200 lines.

Separate the accounts if you run agents in parallel

If you routinely run two or three sessions at once, one subscription is structurally the wrong shape. Two Max 5x accounts cost the same as one Max 20x and give you two independent windows.

Plan limits versus a 429

One more distinction, because the fixes are unrelated. A plan limit is a subscription window closing. A 429 is an API rate limit on a key, a Bedrock project, or a Google Cloud project.

API Error: Request rejected (429) · this may be a temporary capacity issue.
If it persists, check https://status.claude.com.

• Run /status and confirm the active credential is the one you expect. A stray ANTHROPIC_API_KEY in your environment routes requests through a low-tier key instead of your subscription, and this is the single most common cause of a surprise 429.
• Check your provider console for the active tier and request a higher one if needed.
• Reduce concurrency: lower CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY, avoid many parallel subagents, or switch to a smaller model for high-volume scripted runs.

If it is a plan limit and you need to finish today, /usage-credits buys overflow on Pro and Max, or sends a request to your admin on Team and Enterprise. Remember the cache-lifetime cost: once you are on credits, the prompt cache drops from an hour to five minutes.

Questions people ask

Keep reading