Limits & errors

Plan, wallet, concurrency and protocol limits are enforced before execution.

Updated 2026-10-30

Shared input bounds

Native text, messages and schemas share a 128 KiB serialized bound, 128 messages/items, up to eight functions/calls and 16 KiB arguments/results. Requested output is 32–4,096 tokens within the configured model ceiling; Messages requires an explicit maximum.

Response status

400 means invalid or unsupported input; 401 invalid credentials; 403 missing admission or scope; 402 insufficient funds/allowance; 404 unavailable resource/model; 409 conflicting operation; 429 rate/concurrency limits; 502/503 inference/service failure; 504 deadline exhaustion. The native protocol determines the error envelope.

Recovery

Keep the safe request identifier when available. Never share a provider or customer key in a support report. A timeout or network error may leave usage pending; a new independent request may cost again.