Home › Claude API lesson 3: Batches, caching
Claude API lesson 3: Batches, caching
Two features lower your bill: the Message Batches API runs non-urgent requests at 50% off, and prompt caching re-uses a long, repeated prompt instead of paying full price every time.
Step by step
- Group requests that don't need an instant answer and send them to the Message Batches API.
- Poll the batch; most finish in under an hour.
- For a long prompt you reuse (instructions, a document), add one top-level cache_control field to turn on automatic caching.
- Or place cache_control on specific content blocks for exact control.
- Compare cache read and write costs on the pricing page with your normal input cost.
Try it: examples
- Good batch job
Classify each of these 5,000 support tickets as billing, bug or how-to, and give a one-line summary.
Good to know
- Cache reads cost far less than normal input (for example US$0.20 vs US$4 per million tokens on Opus 5.5).
- Default cache lifetime is 5 minutes; a 1-hour option exists.
- Keep the part you cache identical between requests, at the start of the prompt.
Sources (official pages, read Nov 8, 2026)
Written from Anthropic's official documentation and product pages; we don't run hands-on tests. Features and limits vary by plan and country.