The quick read
- Assign stable request identifiers and preserve their relationship with inputs.
- Retry only the items that need it, using a process that will not create duplicate records or overwrite reviewed work unexpectedly.
Choose suitable jobs
Nightly classification or a large archive-tagging task may tolerate delayed completion. An interactive conversation usually cannot. Decide based on the task’s deadline and dependencies, then check the provider’s current batch limits and completion expectations.
Give every item an identity
Assign stable request identifiers and preserve their relationship with inputs. Do not assume results return in submission order. Separate successful outputs from failed items, and record why an item failed. Validate a small sample before submitting a large collection.
Recover selectively
Retry only the items that need it, using a process that will not create duplicate records or overwrite reviewed work unexpectedly. Track accepted, completed and failed counts. For an illustrative document archive, retain the input version with each classification so later corrections remain meaningful. Batch execution is useful when scheduling flexibility improves the workflow. It should still produce an inspectable final inventory, with clear ownership of unresolved items rather than a vague message that the batch mostly succeeded.
Sources & notes
AI-assisted editorial content checked against the linked sources.
developers.openai.com — official reference
Sources reviewed for the September 2026 launch edition.