The Message Batches API lets you send a large group of requests to Claude at once and pay 50% less than for normal, real-time calls. The trade-off is time: results can take up to 24 hours, and there is no fixed time for how quickly any batch finishes. It is worth using when the work does not need an immediate answer and nobody is waiting on the screen, such as overnight document processing, bulk classification or scheduled report generation.
What the Message Batches API does
With a normal API call, you send one request and get the answer back straight away. With batches, you submit many requests together as a single job. Claude processes them in the background, and you collect the results when they are ready. Because the platform can run the work when capacity suits it, the price is lower: 50% cheaper than the standard rate.
The key facts to remember are simple:
- 50% cheaper than standard calls.
- Up to 24 hours to complete.
- No fixed turnaround time. Some batches may finish sooner, but you cannot plan around that.
The deciding question: is anyone waiting?
Almost every batch-or-not decision comes down to one question: is a person or a system waiting for this answer right now? If yes, use a normal call. If the answer can arrive hours later without anyone being affected, batches are a strong candidate.
Workloads that suit batches
- Overnight document processing. Summarising or extracting fields from a large pile of contracts, invoices or reports that arrived during the day.
- Bulk classification and tagging. Sorting thousands of support tickets, survey comments or product descriptions into categories.
- Scheduled reporting. Generating weekly or monthly summaries from stored data, where the output is needed by the next morning rather than the next minute.
- Evaluation runs. Testing a prompt against a large set of sample inputs to compare versions before rolling one out.
- Content backlogs. Drafting or rewriting a large catalogue of text that a human will review later anyway.
Workloads that do not suit batches
- Live chat and customer support. A customer will not wait a day for a reply.
- Interactive assistants. Internal tools where an employee types a question and expects an answer in seconds.
- Time-critical decisions. Anything where a late result has a real cost, such as fraud checks that must clear a transaction.
- Multi-step flows that depend on each answer. If step two needs the result of step one, and a person is following along, the delay compounds.
A useful test: if "up to 24 hours, no fixed turnaround" would cause a complaint or a missed deadline, it is not a batch job.
Message Batches vs prompt caching
Both features reduce cost, so they are often confused. They solve different problems and can be considered separately.
| Feature | Saving | Timing | Best when |
|---|---|---|---|
| Message Batches API | 50% cheaper | Up to 24 hours, no fixed turnaround time | Large volumes of work that can wait |
| Prompt caching | Up to 90% off cached input | Immediate | The same large context is reused in real-time calls |
Batches trade speed for price. Caching keeps speed and lowers the cost of repeated context. Our explainer on what prompt caching is covers the second option in detail.
A simple example
Imagine a Malaysian company that receives several hundred supplier invoices every day. Staff do not need the extracted data until the finance team starts work the next morning. Sending the invoices to Claude in a batch each evening costs half the standard rate, and the results are ready by the time the team arrives. If the same company also runs a live chat assistant for customers, that assistant should use normal calls, because replies have to be immediate.
As with any AI output, plan a review step before results feed into decisions; see how to check if AI output is accurate. Check Anthropic's current documentation for the latest limits and behaviour, as details can change.
Where this appears in Claude certification
For the Claude Certified Architect exam (CCAR-F), batches are tested as a decision, not a coding exercise. You need to recognise when a workload is non-blocking and can wait, and when it is not. The six core concepts place this alongside the other architecture choices, and the glossary has a short definition of the term.
Agmo Studio, a Select Partner of the Anthropic Claude Partner Network, runs exam-preparation workshops through Agmo Academy. Certifications themselves are issued by Anthropic. See the Claude Certified Architect program for details.