Skip to main content

Claude Code usage limit reached: what to do next

You are halfway through a task, Claude Code says you have hit your usage limit, and the reset is hours or days away. This page explains how the limits work, how to see one coming, and the four ways to keep going. The last one, handing the task to another tool or account, is the one that doesn't cost you money or time.

How the limits work​

On a Pro or Max plan, Claude Code shares its allowance with Claude in the browser and the desktop app. Two limits apply at the same time:

LimitWhat it isWhen it resets
Session windowA rolling five-hour window that starts with your first messageFive hours after it started
Weekly limitA seven-day cap on total use, with a separate cap for the largest models on some plansSeven days after the week started

Max plans raise both limits (5x and 20x the Pro allowance), but they don't remove them. Long agentic sessions use far more than chat: every tool call sends the conversation, your CLAUDE.md, and the definitions of every connected MCP tool again. Starting a new conversation doesn't reset either window.

Anthropic adjusts the exact numbers from time to time, so check the Claude help center for the current ones. Codex on a ChatGPT plan works the same way, with its own five-hour and weekly windows.

See it coming​

  • /usage inside Claude Code shows your session and weekly bars and when each one resets. /status gives a quick snapshot.
  • Settings > Usage on claude.ai shows the same bars.
  • A status line can show the percentage all the time. specweave statusline does this, and it is what SpecWeave's automatic handoff reads (below).

Claude Code also warns you when you are close to a limit. That warning is the moment to act, because once the limit is reached the session stops mid-step and anything it hasn't written down is stuck in a transcript no other tool can read.

Your options when the limit hits​

1. Wait for the reset​

Free, and fine if the reset is an hour away. Less fine on a weekly limit with five days to go. Before you stop, ask Claude to write down where it is, or the next session (yours, after the reset) starts from a summary of a summary.

2. Pay for more​

Paid plans can turn on extra usage billed at API rates, or you can point Claude Code at an API key. You keep the same session, at a price. A bigger plan raises the ceiling for next time.

3. Use a smaller model​

Switching to a smaller model with /model can get you past a limit that applies only to the largest models. It doesn't give you a fresh session window, and the smaller model may not finish the hard part of the task.

4. Hand the task to another tool or account​

If you also have Codex, Grok Build, Gemini CLI or a second Claude account, the fastest way to keep working is to continue there. The problem is that none of them can read Claude Code's session. Copying a summary by hand loses your uncommitted edits, the decisions you made, and which tasks were actually finished.

SpecWeave fixes that by keeping the state of the work in git:

specweave handoff --reason "out of usage"

handoff releases your task claims, writes handoff.md with where you stopped and what comes next, and pushes your branch and a snapshot of your uncommitted edits. In the other tool, from the same repository:

specweave pickup

pickup applies your edits and prints the next task with its acceptance criteria, its test command, notes and the project memory. You can also just say "hand off" in Claude Code and "pick up" in Codex; the instruction file SpecWeave installs tells every agent what those words mean.

The step-by-step version is Switch from Claude Code to Codex without losing your place.

Hand off before the limit, automatically​

On your own machine, SpecWeave can watch the limit for you:

specweave auto-handoff on          # hand off at 90% of the five-hour or weekly window
specweave auto-handoff on --at 80  # or pick your own threshold
specweave auto-handoff status

In Claude Code it reads the five-hour and weekly usage from the status line. In Codex it reads the rate limits Codex writes to its session log. When the fullest window crosses the threshold, a Stop hook pauses the agent once per window, has it run specweave handoff, and tells you to say "pick up" in the next tool. Under the threshold it adds nothing to your conversation.

If one long turn jumps straight past 90 percent and Claude Code stops on the limit, a second hook writes the handoff itself, without the model, so your latest edits are still pushed.

Cloud sessions (Claude Code on the web, Projects threads, Codex cloud tasks) have no status line or user hooks. There, the agent hands off when Claude Code warns that the limit is near, or when you say "hand off". Auto-handoff rules has the details for each tool.

Make the allowance last longer​

  • Keep the instruction file short. Everything in CLAUDE.md is sent with every request. SpecWeave's AGENTS.md is about 760 tokens, and CLAUDE.md only imports it. See AGENTS.md vs CLAUDE.md.
  • Disconnect MCP servers you aren't using. Their tool definitions are sent every time.
  • Work from a written spec. A short spec.md with acceptance criteria means fewer rounds of "that's not what I meant".
  • Start a fresh session per task. A long transcript makes every later step more expensive. With the task state in the repository, a new session needs one specweave pickup to catch up.

Common questions​

Does starting a new chat reset the Claude Code limit? No. Both the five-hour window and the weekly limit count all your use, across Claude Code and Claude in the browser.

Can Codex continue my Claude Code session? Not directly; Codex can't read Claude Code's transcript. With SpecWeave, the spec, the task state, a handoff note and your uncommitted edits go through git, and specweave pickup in Codex continues from there.

Does it work with a second Claude account instead of Codex? Yes. Project memory in Claude Code stays with one account, so a second account starts cold without a handoff. specweave pickup gives it the same starting point Codex would get.

Do I need the SpecWeave CLI? For automatic handoff, yes. For a manual one, the self-contained skill is enough: npx vskill i handoff. It needs only git and a shell.

See also​