Claude Code ops claude overloaded

Claude is "overloaded" — what it means and how to get unblocked

Illustration of a browser window showing an Overloaded capacity message with a red alert icon
Illustration — the capacity error as it appears in the browser.
Short answer

An overloaded error means Anthropic's capacity is saturated right now. It is not your account, your network, or your config, and nothing you change locally will fix it. Retry with backoff, drop to a lower-demand model tier, or move the work off-peak.

Last verified 22 Aug 2026 Tested against Claude Opus 5 Re-tested monthly

What is actually happening

Overloaded is a server-side capacity signal, returned when demand exceeds available inference capacity for the model you asked for. It is not rate limiting — rate limits are per-account and tell you that you have used too much. Overloaded tells you the service is saturated, and every other user on that tier is seeing the same thing at the same moment.

That distinction matters because it determines what will help. Nothing account-side changes a capacity signal.

The fix

  1. Retry with exponential backoff, not immediately. Wait 2s, then 4s, then 8s. You should see the request succeed rather than return the same error. Hammering retries makes it worse for everyone, including you.

  2. Drop a model tier for the current task. If you asked for the largest model, a smaller one frequently has headroom when the flagship does not. You should see an immediate success on the alternate model.

  3. Move the job off-peak. Saturation tracks working hours in the US and Europe. Batch work you do not need synchronously into off-peak hours. Track your own error rate by hour before assuming a pattern — it varies.

  4. Split the work into smaller requests. Long single requests hold capacity for longer and are more likely to be shed. Smaller units retry more cheaply.

TODO before publishing: log your own overloaded-error rate by hour for two weeks and replace step 3 with measured data. Do not publish a claim you have not measured — the whole differentiator of this site is that the numbers are real.

What does not work

These are the things people try first, and none of them help — because none of them touch capacity:

Real numbers

TODO: replace with measured data from your own logs — error rate by hour, median retries to success, and whether the lower tier cleared it. Include how you measured.

Related