AI Operations Economics (2/4) — Model Routing: The Cost / Quality / Latency Triangle

AI ์šด์˜ ๊ฒฝ์ œํ•™ (2/4) — ๋ชจ๋ธ ๋ผ์šฐํŒ… ์ „๋žต: ๋น„์šฉ·ํ’ˆ์งˆ·์ง€์—ฐ 3์ถ• ์˜์‚ฌ๊ฒฐ์ •

"The most expensive model" is not the answer — over 80% of tasks can hit the same outcome at 1/10 the cost.

ํ•ต์‹ฌ ์š”์•ฝ

  • Routing is a rule system that matches tasks to models. It is not on-the-fly model selection
  • Three axes: cost / quality / latency — every routing decision is a tradeoff among these
  • Primary sources: Anthropic and OpenAI pricing pages; first-person operational data
  • Starting question: "If this task fails, can we just retry?" If yes, start with a smaller model
  • Real automation: a classifier categorizes the task and the router picks the model. Human ad-hoc selection is not routing

1. Why route at all

LLM unit prices vary widely within a single provider.

Model Input ($/1M tokens) Output ($/1M tokens)
Claude Opus 4.7 $15 $75
Claude Sonnet 4.6 $3 $15
Claude Haiku 4.5 $0.80 $4

Opus is roughly 19× the input price of Haiku. Running a classification task on Opus means a 19× bill.

The right question is "is Haiku enough?" — not "is Opus better?" Picking the cheapest model that is good enough is the heart of routing.

2. The three axes

Every routing decision is a tradeoff across three axes.

2.1 Cost

  • Unit price × token usage.
  • Caching, routing, and short outputs together cut cost by ~95% (parts 1/4 and series A's 5/5).

2.2 Quality

  • Measure as retry frequency. If a small model needs three tries, effective cost is 3×.
  • Quality is not an accuracy threshold — it's a retry function.

2.3 Latency

  • Smaller models are faster. For interactive use cases that need 5-second answers, the largest models are out.
  • Background jobs have zero latency weight.

Conceptual formula:

Effective cost = (rate × tokens) × (1 + retry_rate) + (latency × user_impatience_penalty)

In practice this collapses into a rule table.

3. Task × model mapping

A reasonable starting point. Calibrate with your own measurements.

Task Recommended model Reasoning
Classification / labeling / intent Haiku or local oMLX Simple; retry cost is low
Summarization (≤5K input) Haiku or Sonnet Start with Haiku; promote if quality slips
Standard code change / PR drafting Sonnet Best price-quality balance
Hard debugging / design / multi-step reasoning Opus Retry savings outweigh per-call cost
Background monitoring / polling Haiku or local High frequency, latency-tolerant
Interactive chat Sonnet (streaming) Latency requirement ≤2s

Core rule: start small, measure, promote only when needed. Starting with the largest model is almost always overspend.

4. Multi-provider routing

If you mix Anthropic, OpenAI, and local rather than living in a single ecosystem, cross-provider routing is an additional lever.

Provider selection axes: - Data location: can code or documents leave for an external API? If not, route local. - Model strengths: code → Anthropic tends to lead; certain reasoning tasks → OpenAI o-series (decide via benchmarks plus your own measurements). - Availability: when one provider 5xx's, automatically fall back to another.

Fallback strategy: - Primary: most cost-effective provider. - Secondary: same-class model from another provider. - Tertiary: when both fail, escalate to a stronger model (retry is expensive, but the task must complete).

Measurement: track provider × model × success_rate daily. If fallback fires often for one task, reconsider the primary choice.

5. Classifier-driven automatic routing

Human selection per call is not routing — it's improvisation. Real routing is automated by a classifier.

Minimum classifier: - Input: user request text + metadata (task length, available tools, etc.). - Output: {model: "haiku-4.5", reason: "classification task"}. - Implementation: rule-based (keywords) first → upgrade to a small LLM (Haiku) once measured.

Router shape:

[user request] → [classifier] → [model pick] → [LLM call] → [result]
                     │                              │
                     └─[log: task_type, model, cost, latency]

The classifier itself burns tokens, so use the smallest model or rule logic for it. If the classifier exceeds 10% of call cost, simplify to rules.

6. Measurement — What to look at

Effective metrics for routing impact:

  • Model-call distribution: per task type, which model is used and at what rate.
  • Retry rate: per model. A small model with 30%+ retries is too small for the task.
  • Fallback firing rate: too high → primary choice lacks reliability.
  • Average cost per task: before vs. after classifier deployment.

Empirical rule: classifier deployment usually drops average cost by 50%+ . If not, the classifier's rules are too conservative (it isn't routing enough work to small models).

7. Four ways routing breaks

7.1 Permanent regression to small models

  • "Haiku is fast and cheap, route everything to it."
  • Result: retries explode → effective cost equal or higher.

7.2 Default-to-large

  • "Opus is safe, default everything there."
  • Result: 80% of work is overspend.

7.3 Routing without measurement

  • Rolled out routing but never measured impact.
  • Result: six months later the rules don't match reality and no one notices.

7.4 Too many models

  • 5 providers × 3 models = 15 options.
  • Result: classifier becomes unpredictable, fallback chains explode in complexity, debugging suffers.

8. At a glance

Step Core Signal
1. Task classification Rule-based or classifier task_type distribution
2. Model mapping Start small Retry rate
3. Multi-provider Data / strength / availability Fallback firing rate
4. Automation Classifier + router Average cost per task trend
5. Monitoring provider × model × daily success_rate, latency

Routing is not write-once. Model lineups, prices, and internal task mix change — quarterly review is appropriate.

Next up

Part 3/4: Prompt Caching Guide — 1-hour vs 5-minute Cache. If routing decides the model, caching decides the input tokens. Both levers must work together to complete cost reduction.

References

  • Anthropic, Pricing — claude.com/pricing (verified 2026-05-05).
  • OpenAI, Pricing — platform.openai.com/pricing (verified 2026-05-05).
  • Coding Agents in Practice (5/5) — four cost levers (final entry of series A).
  • Series part 1 — token cost structure and pitfalls.

This is part 2/4 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 »