How metering works
The gateway meters the two content endpoints with per-IP token buckets on both requests and bytes. When a bucket empties, requests get429 with a Retry-After header instead of queueing.
Handling 429s in your client
RespectRetry-After, back off, and treat sustained 429s as a signal to move tiers rather than something to code around. The SDKs already behave this way: rate limiting surfaces as a typed error carrying the retry delay (reading with the SDK).