Cut API Costs with Prompt Caching

Why Caching Matters

Repeated system prompts, long documents and few-shot examples are re-processed on every call. Prompt caching stores them once and reuses them, cutting latency and up to 90% of input cost on repeats.

Practical Setup

response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    system=[{
        "text": LONG_STABLE_PROMPT,
        "cache_control": {"type": "ephemeral"}
    }],
    messages=[{"role": "user", "content": question}]
)

Rules

Put stable content first, changing content last: cache breaks on the first difference. Cache system prompts, docs and examples; never cache the user question. Watch the cache-hit rate in the response usage fields.