Description: What Anthropic's low-cache-hit email means, who needs to act, how the claimed 45% savings estimate works, and how to verify real API savings.
1. A performance review from your chatbot?
An email from Claude Console announces that your prompt cache hit rate is low. Then comes the enticing line: you could potentially save up to 45% of your direct API spend. It sounds like an accusation. Have you been asking questions incorrectly? Is Claude about to charge a 45% stupidity tax?
Neither. This is a cost-optimization suggestion. It concerns how often an API request can reuse a previously processed prompt prefix. It is not a service error, a guaranteed discount, or an extra charge. [1]
2. Imagine onboarding the same employee 100 times
Suppose your app includes twenty pages of permanent instructions with every request. The final question changes: summarize an invoice, classify a message, explain a spreadsheet. The shared instructions do not.
Without a cache hit, that shared text is billed as ordinary input each time. With prompt caching, the API can reuse the previously processed matching prefix for a limited time. It's the difference between giving an employee the handbook once and making them attend orientation before every spreadsheet.
The catch is prefix consistency. If a timestamp, random identifier, or unique document appears before your stable instructions, the cache may miss the reusable portion. Put stable material first and the changing request afterward. [1]
3. Does “45%” mean everyone's bill drops by 45%?
No. The message describes a potential saving estimate for direct API traffic. For illustration, a $100 monthly bill becoming $55 would be a 45% total saving. That is arithmetic, not an audited forecast for another account.
According to Anthropic's pricing documentation checked on November 13, 2026, a five-minute cache write costs 1.25 times the normal input price; a one-hour write costs twice the normal input price. A standard cache read costs 0.1 times ordinary input, with lower read multipliers for certain models. Output charges are a separate part of the bill. [2]
In other words, caching rewards repetition, not wishful thinking. Paying a premium to cache material you never reuse is like buying a season pass to visit a museum once.
4. What if you only use Claude Code?
The email explicitly excludes Claude Code from this estimate because Claude Code manages caching automatically. That means the email alone is not a reason for a Claude Code-only user to rush into manual cache settings.
The relevant workloads are direct Claude API calls from apps, scripts, internal tools, or custom automations. A regular Claude chat subscription is not the same spending category as direct API usage.
5. How to check whether caching helps
First, check whether your account actually has direct API charges. Next, identify a long repeated prefix: system instructions, reference documents, or tool definitions. Anthropic supports automatic caching through a top-level "cache_control": {"type": "ephemeral"} setting. When requests end with unique content, an explicit breakpoint at the end of the stable prefix can be better. [1]
Finally, inspect the API response: cache_read_input_tokens measures reused input, cache_creation_input_tokens measures cache writes, and input_tokens counts other input after the breakpoint. Compare actual spending, not just a rising hit-rate percentage. A default five-minute lifetime works for closely spaced requests; a one-hour option costs more to write. [1][2]
6. The punch line
Claude isn't judging your memory. It's pointing out that your software may be paying to reread the same handbook over and over. The best answer to that email is not panic. It's to check the bill, separate reusable instructions from changing questions, and measure the difference. The robot remembers the handbook; your invoices no longer need to forget it.
Sources
- Checked November 13, 2026. The 45% figure is an email estimate, not measured savings for every user.
References (2)
- Anthropic: Prompt caching platform.claude.com
- Anthropic: Pricing platform.claude.com
