Claude Prompt Caching Break-Even: The Math Behind the Savings
Claude prompt caching breaks even on the second request: a cache write costs 1.25x the base input price, a read costs 0.1x (0.05x on Opus 5.5 and Sonnet 5.5), so a single reuse inside the 5-minute window repays the write premium.
Last updated: November 2026
Key takeaways
- 5-minute cache write: 1.25x the input rate. 1-hour write: 2x. Reads: 0.1x on standard models, 0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1 and Mythos 5.1.
- Break-even is the second request for the 5-minute cache.
- Minimum cacheable prefix is 512 tokens on current models; shorter prompts are processed uncached with no error.
- One changed byte in the prefix invalidates the cache from that level onward.
- A 10,000-token system prompt at 2,000 calls a day on Opus 5.5 costs about $80 a day uncached and about $4 cached.
All multipliers below are from TrueFoundry's prompt caching breakdown; verify against Anthropic's docs before you commit a budget.
How much does Claude prompt caching cost?
Cache breakpoints are free. You pay for three kinds of input: cache writes, cache reads and normal uncached input. On Opus 5.5 with a $4 per 1M base input price, a 5-minute write is $5 per 1M, a 1-hour write is $8, and a cache hit is $0.20.
When does prompt caching break even?
Take a prefix worth 1 unit uncached. Uncached, N requests cost N units. Cached with the 5-minute option, they cost 1.25 + 0.1 x (N - 1) units. The 0.25-unit write premium is repaid on the first read, which saves 0.9 units, so break-even is the second request. With the 0.05x read rate on Opus 5.5 and Sonnet 5.5 the crossover is near 1.26 requests, so the second request still wins.
Here is something the source does not spell out. The 1-hour write costs 2x, so the same algebra gives 2 + 0.1 x (N - 1) < N, which first holds at N = 3. The 1-hour cache only pays if you expect at least two reuses (an illustrative extension of the source's formula). Use it when traffic arrives every 5 to 60 minutes.
What does caching save in dollars?
TrueFoundry's example: a 10,000-token system prompt on Opus 5.5 with 2,000 tickets a day.
| Scenario | Cost per request | Cost per day | |---|---|---| | Uncached | $0.04 | $80 | | Cached read | $0.002 | $4 | | Cache write (when the cache expires) | $0.05 | per write |
That is about a 95% cut on this block, or roughly $76 a day and $2,280 a month if every call hits (illustrative, ignoring write events). With 2,000 tickets a day, traffic arrives about every 43 seconds, so the 5-minute TTL stays warm. For the gateway-side view, see our post on prompt caching and LLM gateway savings.
What invalidates a Claude cache?
A hit needs a 100% identical prefix, byte for byte, up to the cache_control block. The order is tools, then system, then messages, and a change at one level invalidates that level and everything after it.
- Editing a tool definition misses all three levels.
- Changing the system prompt misses system and messages.
- Adding or removing an image misses messages.
- Placing a breakpoint on a per-request value such as a timestamp or the user message writes a fresh cache on every call and never reads one.
Verify with the usage fields: total input equals cachereadinputtokens plus cachecreationinputtokens plus input_tokens. If the first two are zero, nothing was cached. You get up to 4 breakpoints per request.
How do you build caching into your cost model?
- 1Put stable content first: tools, system prompt, reference docs. Put volatile content last.
- 2Check your prefix is above 512 tokens on current models.
- 3Track cache hit rate as a KPI next to cost per task.
- 4Model cached and uncached scenarios side by side in the Calcaas LLM cost calculator, and compare rates on the Anthropic pricing page.
Frequently asked questions
How much cheaper is Claude prompt caching?
A cache read costs 0.1x the input price on standard models and 0.05x on Opus 5.5 and Sonnet 5.5, so a fully cached prefix is 90% to 95% cheaper per read.
How long does the Claude cache last?
The default cache lasts 5 minutes and refreshes at no charge on each read. A 1-hour option costs 2x on the write.
What is the minimum prompt size for caching?
512 tokens on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, and 1,024 or more on older models. Shorter prompts run uncached with no error.
Why is my cache hit rate zero?
The prefix probably changes between calls. A timestamp, reordered tools or an edited system prompt breaks a byte-for-byte match, so check the usage fields for cachereadinput_tokens. The takeaway: caching pays from the second request, so the real work is keeping your prefix stable. Test the savings on your own prompts in the Calcaas LLM cost calculator. Note: place the JSON-LD above inside a <script type="application/ld+json"> tag in the page head.
More from the blog
LLM EconomicsFixed-Price vs Pay-As-You-Go Inference: Breakeven Utilization
Fixed-price inference plans beat pay-as-you-go only once your traffic fills roughly 60 to 80% of the monthly token allowance; the breakeven utilization is u = P / (A x r), the plan price divided by what the same allowance would cost metered.
LLM EconomicsClaude API Pricing per Million Tokens: Every Model, Cache and Batch Rate
Claude API pricing runs from $0.10 per 1M input tokens on Haiku 5.5 to $8 on Fable 5.2, with output always about 5x the input price, cache reads at a fraction of input, and a 50% discount for batch jobs returned within 24 hours.
LLM EconomicsFable 5.2 Cost per Task: Cheaper Tokens, a 31% Higher Bill
Fable 5.2 cut list prices 20% to $8 input and $40 output per 1M tokens, yet SemiAnalysis measured its cost per task rising 31%, from $67.16 to $88.15, because the model uses far more tokens to check its own work.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.