---
title: The stable part first — AI hygiene, lesson 3
description: Keep the long part that does not change first and unchanged, so it is read from the cache at a twentieth of the price on Opus 5.5; one changed byte misses.
url: https://instibrain.ai/guides/ai-hygiene/the-stable-part-first
updated: 2026-11-08
site: https://instibrain.ai/llms.txt
---

lesson 3 of 8 · ai hygiene

# The stable part first.

Keep the long part that does not change at the top, and keep it unchanged. Then the provider reads it from its cache at a fraction of the price.

## new words

**Prompt caching**
The provider keeps the beginning of a prompt it has read, so the next call that starts the same way reads it from the cache.

**Prefix**
The beginning of the prompt, from its first token up to a point.

**Cache hit, cache miss**
A hit: the beginning matched, cheap. A miss: something in it changed, full price, and the cache is written again.

**Cache lifetime**
How long the cache waits for the next call: 5 minutes by default, renewed free at every hit, or 1 hour at a higher write price.

## the lesson

The provider matches the new prompt against the one it cached, from the first token on. A hit needs the beginning to be exactly the same. One changed character near the top, and everything after it is read at full price, then cached again.

| On Opus 5.5, per million tokens | Price | Of a fresh read |
| Fresh read | $4.00 | 1× |
| Cache write, 5-minute lifetime | $5.00 | 1.25× |
| Cache write, 1-hour lifetime | $8.00 | 2× |
| Cache read | $0.20 | 0.05× |

These are Anthropic's list prices. A cache read is a twentieth of a fresh read on Opus 5.5 and Sonnet 5.5, and a tenth on most other Claude models. For a 100,000-token window on Opus 5.5:

```
first message, cache write:  100,000 × $5 / 1,000,000   = $0.50
each message after, a hit:  100,000 × $0.20 / 1,000,000 = $0.02
20 messages: $0.50 + 19 × $0.02 = $0.88, against $8.00
```

The cache lives 5 minutes by default, and each hit starts the 5 minutes again for free. Some tools write the 1-hour cache instead, at twice the price of a fresh read. After a pause longer than the lifetime, the next message writes the cache again: $0.50.

The order is fixed: tools, then the instructions, then the conversation, and a change at one level misses that level and all after it. So keep the instructions and the pages first and unchanged, and put what changes, your question, last. Changing the model or the effort level halfway through also misses: choose them when the window opens.

Claude Code and instibrain keep this order for you: instructions and pages load at the start and stay the same through the session.

One prompt, two next messages, on Opus 5.5: $0.20 per million from the cache, $4 at full price. A change near the top, the line, makes everything after it full price again.

## try it

Type this and press Enter:

```
A window of 100,000 tokens, 20 messages on Opus 5.5. Reading costs $4 per million, a cache write $5 per million, a cache read $0.20 per million. What do the 20 messages cost to read with no cache, and with the cache written once and hit 19 times?
```

## you should see

- $8.00 with no cache.
- $0.88 with the cache: $0.50 for the write, then 19 hits at $0.02.

---

instibrain is made by WebWork AI, Inc. Everyone uses the same models · the difference is what yours knows.
