What Is Prompt Caching?
How to Save 90% on Repeated AI Context

How Prompt Caching Works

Without caching, every API call processes your entire prompt from scratch — even if the first 10,000 tokens are identical across all calls (e.g., the same system prompt or document).

With caching, you mark a prefix as cacheable. The first call writes it to cache (cache write, charged at a small premium). All subsequent calls within the TTL read from cache (cache read, charged at 90–97.5% off the standard input price, depending on the model).

Cache Pricing: All Claude Models

The Break-Even Calculation

Cache writes cost 25% more than standard input. On Haiku, cache reads cost 10% of standard input (5% on Sonnet 5.5 and Opus 5.5, 2.5% on Fable 5.1). Break-even = number of reads where total cost = standard input cost for all calls.

Formula: Break-even reads = 1.25 / (1 - 0.10) = 1.39 reads → break even after 2 reads

After 2 cache reads, every additional read saves at least 90% vs standard pricing. For any system prompt used more than twice, caching is profitable.

Worked Example: SaaS Chatbot with 100K Calls/Month

System prompt: 2,000 tokens (company context, instructions, persona). Each call also includes user message (200 tokens) + history (300 tokens).

When Caching Helps Most

• Large system prompts (500+ tokens): Instructions, persona, product context, examples — anything repeated across all calls
• Document-in-context: The same policy document, knowledge base, or legal text injected into every query
• Agent memory: A persistent memory block representing the agent's accumulated context
• Few-shot examples: A bank of 10–20 labeled examples included in every prompt for consistency

When Caching Does NOT Help

• Dynamic prefixes: If every call has a unique user-specific prefix, there's nothing to cache
• Low-volume endpoints: Below ~3 calls/hour on the same prefix, cache may expire before it's useful
• Short system prompts: Anthropic caches only prefixes of at least 512 tokens on the current models and 4,096 tokens on the legacy Haiku 4.5 — shorter prompts are processed without caching and no error is returned
• Output-dominated costs: If you're generating long outputs (reports, content), the savings on cached input tokens are less impactful relative to output cost

TTL Options: 5-Minute vs 1-Hour Cache

Anthropic offers two cache durations:

• Ephemeral (5 minutes): Write costs 25% premium. Best for high-frequency chatbots where the same user makes multiple turns quickly
• 1-hour cache: Write costs 2× standard input. Best for document Q&A sessions where context needs to persist across a longer working session

For most chatbots, ephemeral (5-min) caching is sufficient — users rarely take more than 5 minutes between messages. Use 1-hour for document analysis workflows.