Reference: the Message Batches API
Claude Certified Architect - Foundations (CCAR-F) › Domain 4: Prompt Engineering & Structured Output
The Message Batches API, from the API documentation
Domain 4 asks when batch processing is appropriate and how to size submissions against an SLA. The guide gives three properties. The documentation gives the operational limits you need to actually run one.
The trade
Message Batches API
- 50% reduction
- most batches finish in under 1 hour
- up to 24 hours
- a batch that does not finish in 24 hours EXPIRES
Size and retention
| Limit | Value |
|---|---|
| Requests per batch | 100,000 |
| Batch size | 256 MB |
| Result availability | 29 days from created_at |
Whichever of the two size limits is reached first applies. Note that the 29 days runs from batch creation, not from when processing ended.
custom_id
Each request carries a unique custom_id and each result carries it back. This is not optional bookkeeping: the documentation states plainly that order is not guaranteed, so position cannot be used to match a result to its request. The same field is what identifies exactly which items failed, which is what makes resubmitting only the failures possible.
Sizing submissions against an SLA
The arithmetic the exam asks for is a worst case. A document arriving just after a submission window closes waits the full submission interval, then takes up to the full processing ceiling:
submission interval + processing ceiling ≤ your SLA
With a 24-hour ceiling and a 30-hour promise, a 4-hour submission window gives 28 hours worst case and 2 hours of margin. Learn the relation rather than the figures — a question can change any of the three numbers.
Sources
- Batch processing — https://platform.claude.com/docs/en/build-with-claude/batch-processing (retrieved 2026-08-30)