Per key, never per IP
An IP-based limit is wrong in both directions: several customers behind one NAT throttle each other, and a caller distributed across many addresses evades it entirely. A key is the unit you control, so it is the unit we limit. Running two workers? Give them a key each and they get a budget each.Headers
Every rate-limited response carries your remaining budget. Read these rather than counting requests yourself — the window slides, so your count and ours will not agree.
Exceeding the limit returns
429 rate_limited. Back off until x-ratelimit-reset, with jitter — retrying the instant the window opens is how a fleet of workers synchronises itself into a thundering herd.
Limits
The default is 120 requests per minute per key. If your workload needs more, ask — the limit exists to protect shared upstream capacity, not to sell you a tier.A rate limit is not a spend control. It caps the rate of requests, not their cost. To bound what a project can spend, set a monthly budget — see Billing.