Claude Code & Codex limits, explained.

The short version

Both tools give you two allowances that run at the same time: a short five-hour window and a weekly one. Whichever runs out first stops you.

How resets work

• Claude: two clocks. The session limit resets every five hours and the weekly limit once a week. The status line docs describe the five-hour window as rolling. Anthropic doesn’t publish when the weekly reset falls, so read it off your own account: the limit message, /usage and Settings › Usage show the reset time.
• Switching models doesn’t reset anything. “You’ve hit your session limit” and “You’ve hit your weekly limit” cover all models. A model-specific message (“You’ve hit your Opus limit”) is different: switching to another model family with /model keeps you working.
• Mid-answer. If you hit the five-hour limit while Claude is replying, Claude Code may finish the current step (“Usage limit reached · wrapping up”). This is small and capped, counts toward your weekly limit, and is still rolling out.
• Waiting it out. Claude Code v2.1.234 and later can wait for the reset and pick up the interrupted task on its own; see /rate-limit-options.
• Codex. Hit either limit and you wait for that reset or buy credits. OpenAI’s own tip: check the usage dashboard every week or two to see your pace.

See where you stand

• /usage in Claude Code. Plan usage bars, plus a breakdown of what used them: skills, subagents, MCP servers, long context, cache misses. Press d or w to switch between 24 hours and 7 days. The breakdown only covers this machine.
• claude.ai › Settings › Usage. Progress bars for the five-hour session and the weekly limit.
• /status in the Codex CLI, or the dashboard at chatgpt.com/codex/settings/usage.
• Always on screen (Claude Code). A status line script receives rate_limits.five_hour.used_percentage and rate_limits.seven_day.used_percentage (Pro and Max, after the first reply in a session). Add this to ~/.claude/settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline.sh"
  }
}

Then save this script as ~/.claude/statusline.sh and run chmod +x ~/.claude/statusline.sh:

#!/bin/bash
# ~/.claude/statusline.sh  (chmod +x it; needs jq)
input=$(cat)
FIVE_H=$(echo "$input" | jq -r '.rate_limits.five_hour.used_percentage // empty')
WEEK=$(echo "$input" | jq -r '.rate_limits.seven_day.used_percentage // empty')
OUT=""
[ -n "$FIVE_H" ] && OUT="5h $(printf '%.0f' "$FIVE_H")%"
[ -n "$WEEK" ] && OUT="${OUT:+$OUT · }7d $(printf '%.0f' "$WEEK")%"
echo "${OUT:-limits: after the first reply}"

What burns it fastest

• Long sessions. Claude Code sends the whole conversation with every request. A one-line question in a session that has been open all day still costs the whole conversation.
• Coming back after a break. On a subscription the prompt cache lasts an hour. The first message after that re-reads your whole context at full price.
• Subagents, teams and loops. Each subagent sends its own requests. Agent teams use roughly 7× the tokens of a normal session when teammates run in plan mode. A /loop keeps firing while you are away.
• Compacting a huge context. /compact reads everything it summarizes, so it is a big request itself. /clear costs nothing.
• Codex extras. Fast mode uses 2.5× the included usage, Ultrafast 8×. Image generation burns limits about 3–5× faster. Every MCP server adds context.

8 habits that stretch them

1. Clear between unrelated tasks. /rename the session, /clear, and /resume it later if you need it.
2. Match the model to the job. Sonnet handles most coding. Keep Opus for hard design calls, and give simple subagents model: haiku.
3. Plan before it codes. Shift+Tab into plan mode for anything big. Press Esc the moment it heads the wrong way; /rewind goes back.
4. Keep CLAUDE.md under 200 lines. It loads every session. Move workflow-only instructions into skills, which load when used.
5. Turn off MCP servers you aren’t using. /mcp lists them. A CLI such as gh costs less context than an MCP server.
6. Lower the effort for easy work. /effort for renames, small fixes and lookups.
7. Come back within the hour, or start fresh. After an hour away, a fresh session with a short summary is cheaper than poking the old one.
8. Keep noisy output out of the main thread. Let a subagent run the test suite or dig through logs and report back, or use a hook that trims test output to the failures.

On Codex, the same habits apply. When you are close to a limit, run work locally instead of as a cloud task, skip Fast and Ultrafast, and check /status before you start something long.

sources

Checked against these official pages on 20 October 2026. Plans and limits change; if a page here disagrees with them, the official page wins.

1. Using Claude Code with your Pro or Max plan · Claude Help Center
2. Understanding usage and length limits · Claude Help Center
3. Usage limit best practices · Claude Help Center
4. Manage costs effectively · Claude Code docs
5. Customize your status line · Claude Code docs
6. Codex pricing and usage limits · OpenAI