Claude Code rate limits: how the windows work, and how to stop losing runs to them

Nothing about an agent run hurts quite like watching it die at 80% because a usage window closed. This is a practical guide to how the limits behave and what actually helps — most of it needs no tools at all.

How the limits behave

Claude subscription usage is metered in usage windows: a session window measured in hours (the one that usually bites mid-run) and a longer weekly ceiling on top. Two things matter about how they behave:

The session window starts at your first prompt and resets fully when it ends — your usage page shows the exact reset time. It floats with when you start working, not with a clock time like midnight.
The moment a window fills, everything stops mid-task: the agent doesn't get a grace period to finish the file it's editing.

Exact quotas vary by plan and change over time, so treat any specific number you read as stale — what doesn't change is the shape: short-window burn rate is what kills long runs.

What actually helps

Start big runs on a fresh window. If you burned most of the short window exploring, don't launch the two-hour migration now. Take the break; start when it reopens.
Make stalls cheap instead of rare. Have the agent commit after every green test run. When a limit (or anything else) kills the session, you resume from the last commit instead of replaying the afternoon.
Watch the window, not the vibes. The pain is never the limit itself — it's being surprised by it. Any visibility beats none: check your plan's usage page before long runs.
Keep a second login if you genuinely need continuity. Plenty of heavy users keep a personal and a work account. The trick is switching before the run stalls, which is exactly the kind of thing a human forgets and a tool doesn't.

DevFob's usage view: Claude on a Max plan at 67% of the 5-hour session window resetting in 1 hour and 82% of the weekly window, and Codex at 38% session and 91% weekly, each with a coloured progress bar and reset time. — Both windows, per provider, with the reset time next to each — the point is seeing 82% before it becomes 100%.

Where DevFob fits (disclosure: this is our product). DevFob's menu bar shows live usage meters for every Claude login you keep — the short window and the weekly one — so you see a wall coming before you hit it. And if you keep more than one account, it can auto-switch to the login with the most headroom the moment a window fills, so the run continues instead of stalling. It also mirrors sessions to your phone and watch so a stalled-on-a-question run never waits for you either. Local-only, end-to-end encrypted, $15 once. 14-day free trial.

Things that don't help

Prompt tricks. The meter counts what the model processes; phrasing doesn't meaningfully change agentic burn.
Rage-retrying. A filled window reopens on its schedule; retrying just queues frustration.
One giant uncommitted session. The longer the run, the more a mid-run stop costs. This is a discipline problem wearing a rate-limit costume.

Try DevFob free — 14 days $15 until Aug 14, then $29 · no account · macOS 14+

Based on how Claude subscription limits behave as of July 2026; specifics vary by plan and change — check Anthropic's docs for current numbers. Something wrong here? Tell us.