Claude Prompt Caching Cost Calculator

Caching Pricing Rates

Cache writes occur once per 5-minute TTL window. All subsequent requests within that window read from cache at 10% of standard input cost.

Real-World Savings: Chatbot Example

A chatbot with a 2,000-token system prompt, 300-token average user message, 200-token average response, 10,000 conversations/day, 3 turns per conversation:

RAG Pipeline Savings Example

A document Q&A app that loads a 20,000-token document context per session, 500 questions/day, 10 turns per session:

How to Enable Prompt Caching

Add cache_control to the last content block you want cached. You can have up to 4 cache breakpoints per request.

Python (system prompt caching)

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system=[{
        "type": "text",
        "text": "You are a helpful assistant...",  # your system prompt
        "cache_control": {"type": "ephemeral"}  # ← add this
    }],
    messages=[{
        "role": "user",
        "content": user_message
    }]
)

# Check cache hit in response
print(response.usage)
# Usage(cache_creation_input_tokens=1247, cache_read_input_tokens=0, ...)
# Second call: cache_read_input_tokens=1247 (90% cheaper)

Document + system prompt caching (RAG)

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system=[{
        "type": "text",
        "text": "Answer questions based on the document below.",
        "cache_control": {"type": "ephemeral"}
    }],
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": document_content,  # large document
                    "cache_control": {"type": "ephemeral"}  # ← cache the doc
                },
                {"type": "text", "text": user_question}
            ]
        }
    ]
)

Caching vs No-Caching Breakeven

Caching is always beneficial when you send the same context more than once within 5 minutes. The write cost is recovered on the second request:

Cache writes cost 125% (1.25×) of standard input. Cache reads cost 10% (0.10×). You break even on the 2nd read — the write overhead of 0.25× is recouped immediately since reads save 90%. Starting from the 2nd request, every cache hit gives you 90% off that token segment.