Prompt Caching with Anthropic: Save 90% on Claude API Costs

Anthropic's prompt caching feature can reduce your Claude API costs by up to 90%. If you're calling Claude repeatedly with the same system prompt, documentation, or codebase context, you're likely paying 10–50x more than you need to.

This tutorial explains how prompt caching works, when it applies, and how Attune AI enables it automatically for your workflows.

What Is Prompt Caching?

Every time you call the Claude API, Anthropic processes your entire input — system prompt, context, and user message — from scratch. For large inputs (codebases, docs, long system prompts), this adds up fast.

Prompt caching lets Anthropic store a processed version of your prompt prefix on their servers for up to 5 minutes. Subsequent requests that start with the same prefix retrieve the cached version instead of reprocessing it.

Cost impact:

• Cache write: ~25% more expensive than a standard input token
• Cache read: ~90% cheaper than a standard input token
• Break-even: After 2 requests with the same prefix, you're saving money

When Does Caching Apply?

Caching applies when:

1. Your prompt prefix exceeds 1,024 tokens (the minimum cacheable size)
2. Consecutive requests share the same prefix (system prompt + context)
3. The cache hasn't expired (5-minute TTL on most models)

Common high-value caching scenarios:

• Multi-turn conversations with a long system prompt
• Code analysis where you send the same codebase to multiple prompts
• Document Q&A where the document is the prefix

How to Enable Prompt Caching in the Anthropic SDK

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": "You are an expert Python developer...",
            "cache_control": {"type": "ephemeral"},  # Mark as cacheable
        }
    ],
    messages=[
        {"role": "user", "content": "Review this function for bugs."}
    ],
)

print(f"Cache read tokens: {response.usage.cache_read_input_tokens}")
print(f"Cache write tokens: {response.usage.cache_creation_input_tokens}")

The cache_control: {"type": "ephemeral"} annotation tells Anthropic to cache everything up to that point in your prompt.

Caching Multi-Turn Conversations

For multi-turn conversations, cache the system prompt and update the message history:

messages = []

for user_input in conversation_turns:
    messages.append({"role": "user", "content": user_input})

    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        system=[
            {
                "type": "text",
                "text": LONG_SYSTEM_PROMPT,
                "cache_control": {"type": "ephemeral"},
            }
        ],
        messages=messages,
    )

    assistant_message = response.content[0].text
    messages.append({"role": "assistant", "content": assistant_message})

The system prompt is cached on the first call and reused on every subsequent call in the conversation.

How Attune AI Handles This Automatically

Attune AI's tier-based model routing enables prompt caching automatically for all workflows. You don't need to annotate your prompts manually.

from attune.workflows import SecurityAuditWorkflow

workflow = SecurityAuditWorkflow()
# Prompt caching is applied automatically
result = await workflow.execute({"path": "src/"})

print(result.cost_savings)  # e.g., "Saved 87% vs uncached"

Behind the scenes, Attune AI:

1. Identifies the cacheable prefix (system prompt + codebase context)
2. Adds cache_control annotations automatically
3. Tracks cache hit/miss ratios in the telemetry dashboard
4. Routes repeated identical prompts to the cache

Measuring Your Savings

Use the Attune AI telemetry CLI to see your caching stats:

attune workflow run security-audit
# Output includes:
# Cache hit rate: 84%
# Tokens saved: 142,300
# Estimated cost saved: $0.43 (this run)

Or check the usage tracker programmatically:

from attune.telemetry import UsageTracker

tracker = UsageTracker()
stats = tracker.get_stats()
print(f"Cache hit rate: {stats.cache_hit_rate:.1%}")
print(f"Total saved: ${stats.total_savings:.2f}")

Summary

Scenario	Without Caching	With Caching	Savings
10-turn conversation, 5K token system prompt	$0.50	$0.07	86%
Codebase scan, 20K token context, 5 prompts	$2.00	$0.22	89%
Document Q&A, 10K token doc, 10 questions	$1.00	$0.13	87%

Prompt caching is one of the highest-leverage optimizations available in the Anthropic API. If you're running any repeated Claude workflows, enable it immediately.

Further Reading

• Anthropic Prompt Caching Documentation
• Attune AI Model Tier Routing
• Semantic Caching for LLMs