AI Operations Economics (1/4) — Token Cost Structure and Measurement Pitfalls

AI ์šด์˜ ๊ฒฝ์ œํ•™ (1/4) — ํ† ํฐ ๋น„์šฉ ๊ตฌ์กฐ์™€ ์‹ค์ธก ํ•จ์ •

"Token rate × usage" looks simple, but the actual bill always diverges from that simple formula. Where it diverges is the starting point of operations.

ํ•ต์‹ฌ ์š”์•ฝ

  • Token cost = (input × input rate) + (output × output rate). Everyone gets that far.
  • Five real-world pitfalls: cache-write multipliers / thinking tokens / system-prompt leakage / tool-definition cost / multi-turn accumulation
  • Primary sources: Anthropic / OpenAI official pricing and caching docs, plus first-person operational measurements
  • Making the bill predictable is the precondition of every cost-reduction lever

1. The simple formula vs. reality

The one-line cost model:

Cost = (input tokens × input rate) + (output tokens × output rate)

Most cost calculators stop here. In production, the bill almost always exceeds the prediction. The gap appears in five places.

2. Pitfall 1 — Cache-write multipliers

The most common misunderstanding. "Cache hit means 1/10 cost" is true, but cache writes are more expensive than baseline.

Operation Anthropic rate multiplier
Cache miss (regular input) 1.0×
Cache write (5-min TTL) 1.25×
Cache write (1-hour TTL) 2.0×
Cache hit 0.1×

Break-even math: - 5-min cache: 0.25× extra cost → 0.9× saving on hit. Break-even after 2 hits. - 1-hour cache: 1.0× extra cost → 0.9× saving. Roughly 2 hits to break even, but the 1.0× upfront is large enough that you should validate hit rate before adopting.

Measurement pitfall: many bills lump cache-write tokens into the same line as base input, which makes it look like input usage spiked. In reality the 1.25×–2× multiplier is the spike.

3. Pitfall 2 — Thinking tokens

Reasoning modes (extended thinking) cause the model to emit tokens to itself before answering. Those tokens still bill at the output rate.

  • Tokens the user never sees on screen still appear on the bill.
  • Harder tasks generate exponentially more thinking tokens.
  • Operating without knowing whether thinking mode is enabled means the bill can balloon unexpectedly.

Measurement rule: when the API response separates usage.output_tokens from usage.thinking_tokens, log the latter independently. Decide which task types actually need thinking per category.

4. Pitfall 3 — System-prompt leakage

The same system prompt on every call adds those tokens to every request's input.

  • A 5K-token system prompt × 100 calls/day = 500K tokens of baseline per day. Without cache, all at full price.
  • The longer the system prompt, the more asymmetric the cost on small tasks. If "classify one sentence" is 1/100 the size of the system prompt, ~99% of the cost belongs to the prompt.

Mitigation: (a) cache the system prompt; (b) split small tasks (classification, summarization) into a lighter system prompt or route to a smaller model.

5. Pitfall 4 — Tool-definition cost

When using MCP / function calling, tool schemas enter the input on every call. More tools, more input bloat.

  • 50 tools × ~100 tokens each = 5K-token fixed cost per call.
  • The user invokes only one or two, but the model has to see the whole catalog to decide.
  • Cache hits help, but a frequently-changing tool catalog breaks the cache.

Mitigation: expose subset catalogs per task type. Disable irrelevant tools per session, or split into separate agents.

6. Pitfall 5 — Multi-turn accumulation

Conversation- and agent-style calls re-send the entire prior message history every turn. The first turn's input is billed 10 times in a 10-turn conversation.

  • Without cache hits, an N-turn conversation's input cost grows O(N²).
  • "Why is the same task twice as expensive as yesterday?" is almost always the conversation got longer.
  • When auto-compaction triggers, all tokens up to the compact point are billed for that day before context resets.

Mitigation: start a new session when a task ends. For long conversations, explicitly truncate context.

7. Measurement — Where to look

Anthropic Console (Usage): daily, weekly, monthly tokens broken down into cache write / cache hit / regular input / output. The single most reliable source.

Claude Code /cost (CLI): current-session tokens and cost. See what a task just cost immediately after finishing.

Custom router logs: when running multiple providers, log a per-call cost record in the router. Aggregate daily by provider × model × task_type.

Empirical signals: - cache_hit_rate < 30% → re-examine cache placement (front of cache block likely changing). - thinking_token_share > 50% → classify whether thinking is really needed. - Average input > 2× average output → suspect system-prompt or tool-catalog leakage.

8. Unit economics — Cost per task

Monthly totals don't help decisions. Decompose by task.

Task type Avg input Avg output Avg cost (Sonnet) Note
Classification / labeling 2K 50 Very low Route to Haiku
Summarization (1K text) 5K 300 Low Lower with system-prompt cache
Code change (one file) 15K 1K Medium 30% cut with cache + short output
Multi-agent workflow 50K+ 5K+ High Largest savings opportunity

This unit reveals where to apply caching, routing, and short-output rules.

9. At a glance

Pitfall Signal First mitigation
Cache-write multiplier Input tokens above expectation Measure hit rate before enabling 1-hour cache
Thinking tokens Output tokens asymmetrically large Classify which tasks actually need thinking
System-prompt leakage Small tasks with high unit cost Cache the system prompt + per-task separation
Tool-definition cost Persistent baseline input Subset catalogs per task
Multi-turn accumulation Same task, N× the cost Session separation + explicit context cuts

Once the bill is predictable, the next step is active reduction — model routing, in part 2/4.

Next up

Part 2/4: Model Routing — The Cost / Quality / Latency Triangle. When unit prices differ by 10×, deciding which task goes where is the next-largest lever.

References

  • Anthropic, Pricing — claude.com/pricing (verified 2026-05-05).
  • Anthropic, Prompt Caching — docs.claude.com/en/docs/build-with-claude/prompt-caching (verified 2026-05-05).
  • Anthropic, Extended Thinking — docs.claude.com/en/docs/build-with-claude/extended-thinking (verified 2026-05-05).
  • OpenAI, Pricing — platform.openai.com/pricing (verified 2026-05-05).

This is part 1/4 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 »