Long-running operations
The job model for asynchronous calls.
Anything that can't finish in ~10 seconds — most generation and rendering calls — is asynchronous. You get back a job immediately instead of waiting:
Poll the job's URL, or wait for its <object>.succeeded / <object>.failed webhook (see
Webhooks) once that ships. status moves through queued → running → succeeded|failed|canceled;
credits_charged fills in once the result is final — that's the number that actually left your balance,
which can be less than credits_reserved for variable-cost operations. Jobs are kept for at least 7 days.
Retrying a timed-out call is safe. If your request to start the job times out after it was already
accepted, retrying with the same Idempotency-Key returns the existing job instead of starting a second
one — you'll never end up with two jobs (and two charges) from one timeout.
On MCP, the tool call itself waits up to 25 seconds for the job to finish before giving up and handing
you the Job object plus a pointer to the matching <product>_get_job tool — so a quick job still feels
synchronous to the agent, and a slow one degrades gracefully instead of hanging.
Next: Idempotency · Credits