Prompt Caching Savings Calculator

If every request re-sends the same system prompt, instructions, and reference context, prompt caching can cut that repeated portion's cost by around 90% on cache hits. See what it does to your monthly bill.

Cached tokens per request (system prompt + context)

Unique tokens per request

Requests per day

Input price / 1M tokens ($)

Cache-read discount % (approx.)

Cache-write premium % (approx.)

Cache hit rate: 90% of requests read the cache (the rest re-write it)

Monthly comparison (30 days)

Scenario	Cached portion	Unique portion	Total / month

Monthly savings

Bill reduction

Effective price / 1M cached tokens

What caches well: anything identical at the start of every request:

System prompts, tool and function definitions, safety instructions
Few-shot examples and style guides
Large reference documents, schemas, or codebases shared across requests
Conversation history in multi-turn chats (each turn re-reads the prefix)

What does not: content that changes per request (the user's message, retrieved chunks that differ every time). The math: without caching every request pays full price for cached + unique tokens. With caching, hits pay (1 - read discount) on the cached portion; misses pay (1 + write premium) to write it. The 90% read discount and 25% write premium are typical published figures but approximate, as of July 2026 - check your provider (cache lifetime, minimum cacheable size, and exact rates vary). Caching pays off fastest with long shared prefixes and high request rates; it also cuts time-to-first-token.