Caching Plus Batching

Definition

The combined cost-optimisation pattern of layering Prompt Caching on top of the Message Batches API to stack a cache discount on top of the 50% batch discount.

Key points

• Prompt caching works in the Message Batches API at the same pricing (0.1× read, 1.25×/2× write).
• Batch API gives a flat 50% cost reduction; caching adds the read/write multipliers on top.
• Use the 1-hour cache TTL with batches — the 5-minute TTL expires before async batch jobs run, so a 5m cache would be lost.
• Batches are not ZDR-eligible; caching is. If ZDR is required, the batching leg disqualifies the combo — see ZDR Eligibility.
• Concrete case (playbook): follow-up synthesis summaries taking 40+ seconds passing 80K+ accumulated tokens → enable prompt caching on the synthesis subagent to cut transfer overhead.

Why it matters for the exam

• Cost-optimisation questions reward stacking both discounts and, critically, choosing the 1-hour TTL for batches. A frequently tested combination.

Common gotchas

• Defaulting to the 5-minute TTL with batches — it expires before the async job completes, wasting the cache write.
• Assuming the combo is ZDR-safe — the batch leg is not ZDR-eligible.

See also

• Prompt Caching
• Message Batches API
• ZDR Eligibility
• Model Lineup And Pricing

Sources

• netlight-study-guide
• docs-tools-agents
• architects-playbook