Anthropic Claude Prompt Caching Cost Savings & Break-Even Calculator
Claude prompt caching requires a 1.25x write multiplier for the default 5-min TTL (or 2x for 1-hr TTL) and offers cache reads at 0.1x base input (with lower rates for specific models like Opus 5.5/Sonnet 5.5 at 0.05x and Fable 5.1/Mythos 5.1 at 0.025x). For a 6,000-token prefix at a $3/M base rate, 5-minute caching is cheaper than uncached input starting on the 2nd request.
Interactive Parameters & Live Verdict
Rule Diagnostics & Recommendations
Engineering Formula & Verification Notes
Calculations use Anthropic's published prompt caching pricing mechanics. Uncached cost equals n * B. Cached cost equals B * [write_multiplier + (n - 1) * read_multiplier], where B represents the base prefix token cost.
Worked Engineering Example
Consider a 6,000-token system prompt and base documents billed at $3.00 per million tokens on Sonnet. Uncached, 10 requests cost 10 * 0.018 = $0.1800. With 5-minute caching, the first write costs 6,000 * ($3.00 * 1.25 / 1M) = $0.0225, and 9 subsequent reads cost 9 * 6,000 * ($3.00 * 0.1 / 1M) = $0.0048, totaling $0.0273.
- Uncached prefix cost for 10 calls: $0.1800
- Cached prefix cost for 10 calls: $0.0273
- Net savings of 84.8% on prefix processing
Common Sizing & Compatibility Mistakes
Prompt cache keys are strictly byte-exact. Changing whitespace, reordering tool definitions, or appending dynamic system time stamps at the front of the prompt invalidates the cache.
- Placing variable content before stable reference documents
- Ignoring agent execution bottlenecks that exceed the 5-minute TTL
- Miscalculating the write premium when selecting 1-hour TTL options
Prompt Caching Scenario Benchmarks ($3/M Base Rate, 6k Tokens)
| Requests (n) | Uncached Cost | Cached Cost (5-Min TTL) | Savings (%) |
|---|---|---|---|
| 1 Request | $0.0180 | $0.0225 | -25.0% |
| 2 Requests | $0.0360 | $0.0243 | 32.5% |
| 5 Requests | $0.0900 | $0.0297 | 67.0% |
| 10 Requests | $0.1800 | $0.0273 | 84.8% |
Frequently Asked Technical Questions
What is the break-even point for 5-minute caching?
For standard models, 5-minute caching becomes cheaper than uncached input on the second request (n=2), as the read savings outweigh the 1.25x write premium.
When should I use the 1-hour TTL instead of the default 5-minute TTL?
Use 1-hour TTL when your agent workflows or multi-turn tasks experience gaps longer than 5 minutes between calls, preventing unexpected cache expiration.
Are cache breakpoints billed separately?
No. Cache breakpoints themselves are free; you are only billed for writes, reads, and ordinary uncached input tokens.