Rate limit: the budget an API gives you per minute and per day
A rate limit is a cap an API provider puts on your usage per time window: requests per minute, tokens per minute, requests per day. When you exceed it, the API answers HTTP 429 until the window resets. Free tiers of hosted models often reset daily at midnight US Pacific time, and their terms usually forbid using extra accounts to get around the limit.
At a glance
Typical units
Requests per minute, tokens per minute, requests per day
When you hit it
HTTP 429, often with a retry-after hint
Free tier resets
Often daily, at midnight US Pacific
Not allowed
Second accounts or keys to get around the limit
Why does the difference between per-minute and per-day matter?
Because the right reaction is opposite. A per-minute limit clears in seconds, so a
short pause and a retry is fine. A daily limit stays until the reset, often
midnight US Pacific time, so every retry before that is wasted time; the task
should move to another provider or a local model at once, and the exhausted lane
should be marked blocked until the reset.
Can I just use a second free account?
Usually not. Providers such as Groq forbid registering extra accounts or
coordinating several to exceed their limits, and Google restricts projects for
quota circumvention. Besides breaking the terms, it risks losing the main
account too. A system with several lanes (other providers, a local model) gets
the same resilience legitimately.
Handle a limit by
Moving the task to another provider or a local model
Waiting for the window when the task can wait
Sending fewer tokens: shorter prompts, a leaner harness
Not by
Rotating to a second account's key (against most terms)