---
title: Workflow Classification — Step 0 of the Cromus 4-Step Framework
description: Workflow Classification maps AI touchpoints to complexity tiers — Mythos/Capybara, Frontier, Balanced, and Lightweight — ensuring the right model handles every task. Step 0 of the Cromus measurement framework.
canonical: https://cromus.ai/workflow-classification
source_html: https://cromus.ai/workflow-classification
---

# Workflow Classification

> **Step 0 of the 4-step AI workflow measurement framework.**

Workflow Classification maps every AI touchpoint in a process to one of four model complexity tiers. It answers the question: *which model tier is appropriate for each task?*

Getting this right before execution is the foundation of cost governance. Mismatched tiers — using a Frontier model for a Lightweight task — are the single largest source of preventable AI workflow waste.

---

## The four model tiers

| Tier | What it handles | Example models |
|------|----------------|----------------|
| **Lightweight** | Simple classification, extraction, routing, formatting | GPT-4o Mini, Claude Haiku, Gemini Flash |
| **Balanced** | Summarization, drafting, multi-step reasoning with moderate complexity | GPT-4o, Claude Sonnet, Gemini Pro |
| **Frontier** | Complex reasoning, code generation, multi-agent coordination | GPT-4.1, Claude Opus, Gemini Ultra |
| **Mythos / Capybara** | Specialized research, scientific reasoning, maximum-quality generation | Top-tier research models, custom fine-tunes |

---

## How classification works in Cromus

When a SOP is compiled into a SKILL.md, Cromus assigns a complexity tier to each step based on:

1. **Task type** — classification vs. generation vs. reasoning vs. multi-step
2. **IO contract** — structured output vs. freeform generation, input token budget
3. **Failure tolerance** — whether retries and fallbacks are declared
4. **Context dependency** — whether the step depends on long-context accumulation

The classification is **deterministic** — the same SOP always produces the same tier assignments.

---

## Why tier mismatches are expensive

A Frontier model costs 10–50x more than a Lightweight model per token. If 60% of workflow steps are simple extractions or routing decisions running on Frontier models, the bulk of workflow cost is preventable waste — Croms™.

Cromus surfaces these mismatches as concrete optimization actions with dollar estimates:
- "Downgrade 3 classification steps from GPT-4o to GPT-4o Mini → saves $0.042/run"
- "This routing step requires no reasoning capability — any Lightweight model handles it"

---

## Part of the 4-step framework

Workflow Classification is Step 0. The framework:

| Step | Name | What it produces |
|------|------|-----------------|
| **0** | Workflow Classification | Tier map of AI touchpoints |
| **1** | Baseline Cost per Workflow | Fully-loaded per-run cost |
| **2** | Croms Analysis | Preventable waste score |
| **3** | Total Cost of Workflow Ownership (TCWO) | True per-outcome cost over time |

---

## Related pages

- [Baseline Cost per Workflow →](/baseline-cost-per-workflow)
- [Croms™ — preventable AI workflow waste →](/croms)
- [Total Cost of Workflow Ownership →](/total-cost-of-workflow-ownership)
- [Pre-Execution Governance →](/pre-execution-governance)