089 Off De809bae

Prompt Caching Quick Reference

90% off for repeated system prompts

When to Use

✅ Good for:

• Long system prompts (>1K tokens)
• Repeated instructions
• RAG with large context
• Multi-turn conversations

❌ Not for:

• Prompts < 1,024 tokens
• Frequently changing prompts
• One-off requests

Pricing

Minimum Token Requirements

Cache TTL (Time-To-Live)

Note: Cache refreshes on each use. If you call every 4 minutes, it stays alive indefinitely.

Basic Usage

response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": "Your long system prompt here (must be >1024 tokens)...",
            "cache_control": {"type": "ephemeral"}  # ← Enable caching!
        }
    ],
    messages=[{"role": "user", "content": "User question"}]
)

Cache Behavior

Request 1: Cache MISS → Write cache (+25%)
Request 2: Cache HIT  → Read cache (-90%) ✅
Request 3: Cache HIT  → Read cache (-90%) ✅
...
Request N: Cache HIT  → Read cache (-90%) ✅

Important Notes

1. Order matters - Cached content must be at the beginning
2. Exact match - Any change invalidates the cache
3. Images can't be cached - Only text content
4. Multi-turn caching - Use cache_control on conversation history

Source: Anthropic Prompt Caching Docs