The batch discount needs a deadline you can afford
A 50% batch discount is a price condition, and the workflow also needs a completion window, a way to spot missing results and time to recover them.
Overnight tagging looks like an obvious candidate for batch processing. Nobody is watching a chat window, the requests are independent and the results can arrive together. Before moving it, write down the actual deadline. “Ready for tomorrow” is too vague if the next system starts at six in the morning.
Anthropic's Message Batches guide, read on 14 November 2026, offers a 50% discount and says batches expire after 24 hours. A job submitted at ten in the evening therefore needs more planning than the word “overnight” suggests. A completion window extending into the following evening does not promise results by breakfast.
What recovery does to the price
Take 1,000 independent jobs. Assume an ordinary request costs $0.02 and the batch version costs $0.01. These are invented rates. Assume too that an expired request is unbilled and can be completed with one ordinary request. Anthropic documents unbilled expiries; the number of expiries and successful recoveries below are assumptions.
| Assumed expiries | Batch completions | Bill after fallback | Saving against $20 ordinary run |
|---|---|---|---|
| 0 | 1000 | $10.00 | 50% |
| 100 | 900 | $11.00 | 45% |
| 300 | 700 | $13.00 | 35% |
With 100 expiries the final bill is $11, a 45% saving against the $20 ordinary run. The arithmetic still looks good. The deadline may not: if recovery only starts once a request expires, those requests begin after the batch's completion window. Whether that is acceptable is a decision to make before choosing the route.
One finished batch can hold several outcomes
The provider distinguishes succeeded, errored, cancelled and expired requests, and each outcome deserves its own count. A finished batch means processing has ended, not that every document has a usable answer.
A manifest of the jobs you submitted lets you match results back to it. The guide uses a unique custom_id and says results are not ordered, so the join has to be on that identifier. Attaching the first answer to the first document just because both files have the same number of lines invites mismatches.
Our suggested reconciliation is simple: each submitted identifier gets one recorded final disposition. A successful response then goes through the task's usual acceptance check. A format failure is different from a request that never ran, and the recovery policy should keep them apart.
A second attempt needs time
Writing the deadline backwards helps. Leave time to download and reconcile results, validate them and recover missing work. If the provider's processing window leaves no room for those steps, submit earlier or use the ordinary route for the deadline-sensitive part.
Starting a fallback while the batch is still processing means planning for both attempts to finish. The job identifier can be used to prevent duplicate downstream writes, but only if the step that receives the results enforces it, and both bills count where applicable. The table assumes recovery only after unbilled expiry, so it would understate a strategy that pays for two completed copies.
A representative small batch is worth trying before moving the back catalogue, to check the request shape and inspect the returned usage and outcome fields. Elapsed time for that run is an observation, not a service guarantee.
Cache savings stay out of the first estimate
The batch guide says caching can combine with batch pricing, but cache hits are best effort. Start the budget with the batch rate alone and add a cache saving after you have measured reads and writes for the request pattern. Otherwise two attractive discounts can produce a forecast that neither feature promises.
A row in your own results log should name the batch route, submitted count, accepted count, final spend, submission time and deadline, along with outcome counts and the recovery rule. That shows whether the discount bought useful work in time, not just cheaper requests.