Fix "API Error: 503 overloaded_error" and "Repeated 529 Overloaded errors"

Other forms of the same error

• Upstream overloaded after 3 attempt(s): Our servers are currently overloaded. Please try again later.
• API Error: Repeated 529 Overloaded errors
• {"type":"error","error":{"type":"api_error","message":"Upstream returned invalid response"}}

Where this 503 comes from

CCSub sits in front of several upstream lines. When a request fails, it doesn't hand the error straight to you — it tries to recover first:

• On a 5xx, a timeout, or an "overloaded" response, it moves the request to another account or line for the same model and retries.
• Some upstreams signal overload by returning 200 and then putting an error inside the stream. CCSub detects that pattern and fails over the same way.
• Lines that keep failing are temporarily marked down so later requests route around them.

You only get a 503 after those retries are used up and there's genuinely no line available at that moment. In other words, it means "this model is short on capacity right now", and it usually clears within minutes.

What each message means

• 上游线路暂时不可用，请稍后重试（不是您账户余额的问题） — "Upstream temporarily unavailable, please retry shortly (this is not about your account balance)." The English translation is included in the message itself. An upstream line's account had a problem, such as its own balance, and CCSub replaced the raw error with this one.
• Upstream overloaded after N attempt(s): ... — the upstream reported overload, and CCSub tried N accounts that were all overloaded.
• API Error: Repeated 529 Overloaded errors — Claude Code's own message after receiving several 529s in a row and exhausting its retries.
• 502 Upstream returned invalid response and 500 Internal proxy error — the upstream sent back something unparseable, or something unexpected happened while handling the request. Treat these exactly like a 503 and retry.

None of these are billed. Only requests that complete successfully are charged.

The fastest way to get going again

1. Wait a minute or two, then type "continue". The Claude Code docs recommend the same thing: your original message is still in the conversation, so you don't need to paste a long prompt again.
2. Switch models. Use /model, or change the model ID in Codex, Cursor, or whichever tool you use. Different models can run on different lines, so when one model has nothing free, another usually does.
3. Break long tasks up. If one very long turn fails, the whole turn has to be redone. Smaller steps make each failure cheaper.
4. Back off in your own code. On 502, 503 or 529, wait a few seconds, double the delay after each failure, and give up after three to five attempts with a clear message to the user. Retrying in a tight loop will also hit the 429 rate limit.

When to contact support

If the same model returns 503 for more than ten minutes straight, or every model fails, email support@ccsub.net and include:

• The time it happened, to the minute.
• Your tool and version (claude --version or codex --version).
• The model ID and the full error text.
• Your account email — never your API key.

With the time and model, support can find the exact failed request in the logs.

FAQ

Related errors

• API Error: 504 {"type":"error","error":{"type":"timeout_error","message":"Upstream timed out"}}The upstream took too long to start answering. Use streaming (stream: true) in your own code, /compact very long contexts, and split long outputs. A one-off 504 just needs a retry and isn't billed.
• API Error: 402 {"type":"error","error":{"type":"insufficient_balance","message":"余额不足，请前往 https://www.ccsub.net/purchase 充值"}}Your available balance or today's subscription quota ran out. Top up, wait for the daily reset at midnight Beijing time, or — if you do have balance — retry in a few minutes.
• API Error: 429 {"type":"error","error":{"type":"rate_limit_error","message":"API Key request rate exceeded"}}Request-rate errors: raise the key's requests-per-minute or split traffic across keys. Concurrency errors: run fewer sessions or subagents at once — the limit is per account, so extra keys don't help.