Prompt cache breakpoint on text that changes every call

Prompt caching on text that changes every call, so you pay to write and never read.

Rule cache-write-churnCost audit

Why it costs money

With Anthropic prompt caching, writing a prefix to the cache costs more than normal input (1.25x for the 5-minute cache, 2x for the 1-hour cache), and reading it back costs 0.1x. The saving comes entirely from reads.

A cache hit requires the prefix to be identical. If the cached block includes a timestamp, a request id, retrieved documents or the user's message, the prefix is different every time. Every request pays the write premium and none of them get a read, so caching costs more than not caching at all. You can see it in usage: cache_creation_input_tokens on every call and cache_read_input_tokens always zero.

Example

import Anthropic from "@anthropic-ai/sdk";

const anthropic = new Anthropic();

export async function answer(question: string) {
  return anthropic.messages.create({
    model: "claude-sonnet-4-5",
    max_tokens: 1024,
    system: [
      {
        type: "text",
        text: STYLE_GUIDE + "Current time: " + new Date().toISOString(),
        cache_control: { type: "ephemeral" },
      },
    ],
    messages: [{ role: "user", content: question }],
  });
}

import Anthropic from "@anthropic-ai/sdk";

const anthropic = new Anthropic();

export async function answer(question: string) {
  return anthropic.messages.create({
    model: "claude-sonnet-4-5",
    max_tokens: 1024,
    system: [
      { type: "text", text: STYLE_GUIDE, cache_control: { type: "ephemeral" } },
      { type: "text", text: "Current time: " + new Date().toISOString() },
    ],
    messages: [{ role: "user", content: question }],
  });
}

How to fix it

Put the cache breakpoint at the end of the part of the prompt that is identical across requests, such as instructions, a style guide or tool definitions. Move anything that varies (the time, retrieved documents, the user's input) after the breakpoint, where it is billed as normal input.

Then check a few responses: cache_read_input_tokens should be non-zero from the second request on. If it stays at zero, something inside the cached prefix is still changing.

What the audit detects

• Lines that set cache_control, excluding log statements and comments.
• The few lines just above the breakpoint are checked for per-request content: template interpolation, f-strings, Date.now(), new Date(), datetime, timestamps, uuid, random values or a user-role message. It also follows the cached variable within the same file to see whether it is built from retrieved context, chunks, documents or conversation messages.
• Breakpoints on tool definitions or schemas are skipped, and the breakpoint must sit on a message or text block.

Research

Anthropic docs: Prompt caching. Cache writes cost 1.25x to 2x the base input price and reads cost 0.1x. The breakpoint belongs on the last block whose prefix is identical across requests.