Guide
Claude Code “Model Overloaded”: What 529 Means
A “model overloaded” message in Claude Code is HTTP 529. It means Anthropic's capacity for that model is saturated at that moment. It is not your usage limit, it does not count against your quota, and no amount of changing your own settings will clear it — which is worth knowing before you spend an hour reconfiguring things that were never the problem.
529 Is Not 429 — Check Which One You Have
These get conflated constantly, and the fix is different for each.
| 529 Overloaded | 429 Rate Limited | |
|---|---|---|
| Cause | Anthropic's servers are at capacity | You exceeded your own allowance |
| Whose problem | Provider-side, affects many users at once | Account-side, specific to you |
| Counts against quota | No | Yes — it is the quota |
| Does upgrading help | No — paid plans hit 529 too | Often yes |
| What actually helps | Wait, or switch to a model with spare capacity | Slow down, or raise your limit |
Max and Pro subscribers see 529s during busy periods. If you are on a paid plan and getting overload errors, your subscription is not the issue and upgrading will not resolve it.
What To Do, In Order
- Check the status page. Visit status.claude.com first. If there is a live capacity incident, the answer is to wait, and everything below is wasted effort.
- Confirm which path you are actually on. Run
/statusin Claude Code and check your environment variables. If a proxy or a custom base URL is configured, the error may be coming from somewhere other than where you assume. - Switch models. Capacity is tracked per model. Run
/modeland move to a different one — often a smaller or older-generation model has headroom while the newest flagship is saturated. This is usually the fastest way back to working. - Stop retrying past a point. If you have retried the same model four times in a minute and three came back 529, stop. Each further retry queues behind every other client doing exactly the same thing, and makes the congestion worse for everyone including you.
- Come back later. Overload correlates strongly with peak hours. Sessions that fail repeatedly at midday often run cleanly a few hours either side.
The Fallback Flag
Claude Code ships a --fallback-model CLI flag, documented as switching automatically when the default model is overloaded or unavailable. For long unattended runs it prevents a single 529 from ending the session.
Use it deliberately rather than by default. A fallback silently changes which model is answering, which matters when you care about output quality or consistency, and it can make a session hard to reason about after the fact. For interactive work, switching models yourself with /model keeps you in control of what actually ran.
Why This Happens
Frontier models are capacity-constrained. Demand is spiky, and it concentrates on the newest model the week it launches and during working hours in the largest markets. Anthropic sheds load with 529 rather than degrading everyone's experience or queueing indefinitely. It is a deliberate response to genuine saturation, so the timing tends to be correlated across users rather than random.
That also explains why 529s cluster. If you get one, others are getting them at the same moment, and a burst of automatic retries from every affected client is precisely what makes the window longer.
If an Always-On Assistant Is Getting 529s
The calculus differs for an unattended bot. A person can wait; an assistant answering a group chat cannot, and a failed turn is visible to whoever asked.
Two things help. The first is model choice: an assistant handling routine questions rarely needs the newest flagship, and mid-tier models are both cheaper and less prone to capacity contention. The second is provider diversity — if every turn depends on one provider's busiest model, that provider's bad afternoon is your bad afternoon.
On OpenClaw Launch you can change your model from the dashboard without redeploying, and pick from 200+ models across providers, so moving off a congested one takes seconds. See the models directory for what is available.
FAQ
Does a 529 use up my Claude usage limit?
No. An overload error means the request was never served, so it does not consume quota. It is not a penalty and there is nothing to appeal.
Will upgrading to Max stop the overload errors?
No. 529s are capacity-side and Max subscribers report them during busy periods too. Upgrading raises your own limits, which addresses 429s rather than 529s.
How long do overload periods last?
Usually minutes. Occasionally longer during a major launch or an incident. The status page is the only reliable signal — if nothing is posted there and the errors persist for you alone, look at your own network path or proxy configuration.
Is switching models safe mid-task?
Yes for most work. Your conversation context carries over. Expect somewhat different style and reasoning depth, so for a task where consistency matters it can be better to wait than to finish on a different model.