The original title is "Prompt Caching, Batches API, and Model Routing to Cut LLM Costs"

Short summary

Three concrete levers for cutting Claude API costs without switching providers: prompt caching (cache reads at ~0.1x input price), the Batches API (50% discount for async workloads), and model routing (send easy tasks to Haiku, escalate to Opus only when needed). The article includes Python SDK examples, break-even math for cache TTLs, and debugging tips for silent caching failures. Discounts stack — a batch of shared-prompt calls gets both the batch and cache-read discounts.

• •Prompt caching cuts repeated prefix costs to ~0.1x base input price
• •Batches API offers 50% discount for async, non-latency-sensitive workloads
• •Model routing sends 80% of tasks to Haiku, escalating only complex cases to Opus

Generated with AI, which can make mistakes.