Prompt Caching Calculator

Anthropic's Claude prompt caching stores a processed version of your prompt prefix in memory. If you send the same prefix (at least 1,024 tokens for Claude 3 Haiku, 2,048 for Sonnet/Opus) on subsequent requests, cached tokens are charged at 10% of the normal input rate instead of 100%. Cache entries persist for 5 minutes with a sliding window refresh.

Yes. OpenAI automatically caches prompt prefixes for GPT-4o and GPT-4o-mini on prompts of 1,024+ tokens. Cached tokens are charged at 50% of the input rate (not as aggressive as Anthropic's 10%). Cache hits occur when the prefix matches a previous request exactly. No explicit cache management is needed.

Cache hit rate is the percentage of your requests that successfully reuse a cached prompt. A 90% cache hit rate means 9 out of 10 requests use the cached version. Higher hit rates occur when you have a consistent large system prompt sent on every request. Hit rates drop when prompts vary significantly between requests.

Structure prompts with the large, stable content first (system instructions, documents, tools) followed by the dynamic content (user message). Cache is keyed on the exact prefix — even a single character change invalidates it. Pin your static system prompt and append dynamic user input at the end. Use cache_control: {type: 'ephemeral'} in Claude to explicitly mark cache breakpoints.

Caching saves money on repeated large system prompts but adds a cache write cost (25% on Anthropic, 100% on OpenAI — cache writes are charged normally). The savings are largest when: (1) your system prompt is very large (1K+ tokens), (2) cache hit rate is high (>50%), and (3) you make many requests per session. Low-volume or highly variable prompts won't benefit much.