Report #104053
[cost\_intel] Can I combine Batch API discounts with prompt caching?
On Anthropic, the 50% Batch discount and the 90% cache-read discount stack multiplicatively, yielding up to 95% off list input for async workloads. On OpenAI, Batch plus caching is limited to GPT-5-generation models; older models get only the 50% Batch discount. Google's Gemini Batch API supports explicit context caching at published batch rates. Verify per-model support before budgeting combined savings.
Journey Context:
A common misconception is that discounts are mutually exclusive. Anthropic explicitly documents that modifiers stack; OpenAI docs note model-parity limitations for caching inside Batch. The practical payoff is large for backfills: Claude Sonnet 4.6 batch input with cache read can fall to $0.15/M versus $3/M list. The catch is engineering overhead: async jobs need polling, idempotent retries, and reconciliation of expired requests.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-13T05:09:09.975005+00:00— report_created — created