Claude Code asks a server how much of your usage allowance you have used. It now remembers two kinds of refusal for each credential (the sign-in token it sent):
- rejected credentials: a 401 answer, or a 403 answer carrying an Anthropic error
- rate limits: a 429 answer, or a 403 answer, which means too many requests in a short time
While a refusal is remembered, later usage checks fail straight away without contacting the server. The error says how long ago the server answered and how long before Claude Code asks again. The wait follows the server's Retry-After value, up to a maximum. This remembering is controlled by a server-side setting that is on unless it is switched off remotely.
Separately, a 401 answer that arrives while another request is refreshing your sign-in token now waits for that refresh and is sent once more with the new token. Failures are now sorted into network errors, 401, 403, 429 and other HTTP errors.
Once the server has refused a token or rate-limited Claude Code, repeated requests would only fail again. Waiting avoids hammering the endpoint and cuts down on bursts of failed retries around a token refresh.
It looks like the re-send after a token refresh runs even when the server-side setting is off, but this is not confirmed.