AI Operations Economics (3/4) — Prompt Caching Guide: 1-hour vs 5-minute Cache

AI ์šด์˜ ๊ฒฝ์ œํ•™ (3/4) — ์ปจํ…์ŠคํŠธ ์บ์‹ฑ ๊ฐ€์ด๋“œ: 1์‹œ๊ฐ„ vs 5๋ถ„, ์–ด๋””์— ์“ฐ๋‚˜

Caching is not always savings. It is savings if the hit rate is high enough — otherwise it is a loss.

ํ•ต์‹ฌ ์š”์•ฝ

  • Cache hit = 0.1× base input; cache write = 1.25× (5-min) or 2× (1-hour)
  • Break-even: 5-min cache pays back at 2 hits; 1-hour cache around 2–3 hits
  • Primary sources: Anthropic prompt caching docs, OpenAI prompt caching docs
  • Cache targets: system prompts / CLAUDE.md / tool catalogs / frequent static context
  • Decision: 1-hour is not always better than 5-min — only superior when hit rate is reliably high

1. How caching works

LLM caches store the prefix of a prompt. On a later call, if the same prefix appears, computation is reused.

Core rules: - Prefix matching only. One byte changed in the front of the cache block → miss. - Block boundaries must be marked explicitly (cache_control in Anthropic). OpenAI handles it automatically. - Blocks are typically efficient at 1024+ tokens. Smaller blocks have too much overhead.

Hit-time pricing: - Anthropic: base × 0.1 (~10%). - OpenAI: base × 0.5 (~50%, varies by model).

Anthropic is aggressively cheaper, but hit-rate management requires more attention than OpenAI.

2. TTL — 5-minute vs 1-hour

Anthropic offers two TTL options.

TTL Cache write rate Cache hit rate Where to use
5-min (default) base × 1.25 base × 0.1 Within-session repetition; interactive
1-hour base × 2.0 base × 0.1 Long sessions, batch, background

TTL selection rule: - Same context re-called within 5 minutes → 5-min cache. - Re-called between 5–60 minutes → 1-hour cache (if break-even holds). - Gaps over 1 hour → cache expires, and caching becomes a loss.

Claude Code note: Claude Code is known to use 1-hour TTL by default. That means every cache write is billed at 2× base. Bake the 2× multiplier into cost estimates.

3. Break-even math

5-minute cache: - First write: base × 1.25 (extra 0.25× over normal). - Each hit: base × 0.1 → savings of 0.9× per hit. - Break-even: 0.25 = 0.9 × N → N ≈ 0.28 hits. Pays back from the first hit. - Practical rule: enable when 2+ hits are likely.

1-hour cache: - First write: base × 2.0 (extra 1.0× over normal). - Each hit: base × 0.1 → 0.9× savings. - Break-even: 1.0 = 0.9 × N → N ≈ 1.12 hits. Pays back from the second hit. - Practical rule: enable when 4+ hits are likely (safety margin).

Implication: with less than one hit, caching is a loss. "Caching = savings" applied without measurement can raise the bill.

4. What to cache

4.1 System prompts

  • Appear on nearly every call. First-priority cache target.
  • Highest possible hit rate.

4.2 CLAUDE.md / rule files

  • Fixed at the project level — identical across all calls in a project.
  • 1-hour cache fits long sessions.

4.3 Tool catalogs

  • MCP server defs, function-calling schemas.
  • A frequently changing catalog breaks the cache — apply caching after stabilization.

4.4 Frequent static context

  • Design docs, data dictionaries, code style guides.
  • Push changing parts to the back, keep static parts at the front under cache.

Don't cache: - Per-user / per-session data. - Per-call timestamps and random IDs. - Very small contexts (<1K tokens). Cache overhead absorbs the savings.

5. Anthropic vs. OpenAI

Item Anthropic OpenAI
Activation Explicit (cache_control) Automatic
Hit rate base × 0.1 base × 0.5 (model-dependent)
Write multiplier 1.25× (5-min), 2.0× (1-hour) No additional cost
TTL 5-min / 1-hour Per model (typically 5–10 min)
Block boundary Marked explicitly Automatic (1024+ token prefix)

Choosing: - Anthropic gives a deeper discount when you can fully control hit rate. - OpenAI's automatic caching is safer for volatile workloads.

6. Debugging cache misses

If caching seems inactive, check:

  • Front of block changed: does the block start differ each call? (Timestamps in the front are a frequent culprit.)
  • Block size: <1024 tokens disables caching.
  • TTL expired: did calls space out beyond TTL?
  • Reorder: cache misses if cache-block order changed. Verify whether tool definitions auto-reorder.
  • Model swap: cache shares only within the same model.

Metric: compare usage.cache_read_input_tokens vs usage.cache_creation_input_tokens in the API response. Hit rate = read / (read + creation).

7. Operational pattern — Cache-friendly prompt layout

[System prompt — static, 5K tokens]      ← cached
[CLAUDE.md — static, 3K tokens]           ← cached
[Tool catalog — static, 2K tokens]        ← cached
─────────────────────────── cache boundary ───
[User input — varies]                     ← not cached

Core principles: - Put unchanging content first. - Push everything that varies last. - Align cache boundaries with block boundaries.

With this layout, only user input is processed fresh, and the prior 10K tokens enter at 0.1× rate. You buy 10K tokens at the price of 1K.

8. At a glance

Cache type Break-even hits Best workload
5-min TTL 1 (practical: 2) Interactive, short sessions
1-hour TTL 2 (practical: 4) Long sessions, batch, background
Cache Don't cache
System prompts / CLAUDE.md / tool catalogs Per-user data, timestamps, varying random
Frequent static documents Tiny contexts under 1K tokens

Core rule: don't enable caching without measurement. Cache hit rate < 30% → revisit immediately.

Next up

Part 4/4: Context Management Patterns — auto-compact / Memory / RAG Cost Comparison. Caching makes the same context cheaper to send. The next part is about making context smaller in the first place.

References

  • Anthropic, Prompt Caching — docs.claude.com/en/docs/build-with-claude/prompt-caching (verified 2026-05-05).
  • Anthropic, 1-Hour Cache TTL — docs.claude.com/en/docs/build-with-claude/prompt-caching#cache-duration (verified 2026-05-05).
  • OpenAI, Prompt Caching — platform.openai.com/docs/guides/prompt-caching (verified 2026-05-05).
  • Series parts 1 (token cost structure) and 2 (routing).

This is part 3/4 of the AI Operations Economics series.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »