Part 9 — Pick the Right Model Every Time: The Three-Tier Selection Framework

AI Workflow · Module 9

Three-Tier Model Selection

"There is no best AI model. There is only the right model for the job."

Titans Max reasoning · High cost

Workhorses Balanced · Daily use

Sprinters Fast · Low cost · Routine

Most developers pick an AI model when they first set up their tools — and then never change it. They use the same model for formatting fixes and for architecting database schemas. The same model for generating JSDoc comments and for reviewing security-critical authentication code.

This is like choosing between a scalpel and a sledgehammer and then using one for everything. Both are tools. Neither is universally right.

The choice matters because of the three-way trade-off that defines every AI model: reasoning depth, response speed, and cost. Optimizing all three at once is impossible. Choosing which ones to optimize for each task is the skill.

The Three Tiers

🏆 Titans — Maximum Reasoning

When accuracy matters more than speed or cost

High Cost

Slow · Deep

Model	Provider	Strengths
Claude Opus 4.6	Anthropic	Long-context reasoning, nuanced analysis, security review
GPT-4o	OpenAI	Multimodal, strong code generation, broad capability
Gemini 2.5 Pro	Google	1M+ context window, deep reasoning on large codebases

System architecture Security review Complex debugging Novel algorithm design Large codebase understanding

⚡ Workhorses — Balanced Performance

The right choice for 70% of daily development work

Medium Cost

Fast · Accurate

Model	Provider	Strengths
Claude Sonnet 4.6	Anthropic	Fast, strong code quality, excellent instruction following
Gemini 2.0 Flash	Google	Extremely fast, good for real-time interactive coding

API endpoints Feature implementation Standard refactoring UI components Test generation

🏃 Sprinters — Maximum Speed

Fast, cheap, pattern-based — ideal for repetitive and routine tasks

Low Cost

Instant · Pattern-based

Model	Provider	Strengths
Claude Haiku 4.5	Anthropic	Ultra-fast, excellent at pattern tasks, very low cost
Gemini 2.0 Flash Lite	Google	Near-instant responses, ideal for autocomplete and formatting

JSDoc generation Type conversion Boilerplate CSS formatting Simple transformations

The Decision Framework: 2 Questions

Before picking a model, answer these two questions:

Question 1: Task Complexity

Simple / Formulaic → Sprinter
Repetitive, known patterns, no reasoning required

Moderate / Standard → Workhorse
Standard dev work, logical reasoning needed

Complex / Novel → Titan
Multi-system reasoning, novel problems, no clear pattern

Question 2: Business Stakes

Low Stakes → Sprinter OK
Bug here = inconvenience, easily corrected

Medium Stakes → Workhorse
Bug here = user impact, needs proper fix

High Stakes / Critical → Always Titan
Security, payments, data integrity — regardless of perceived simplicity

The high stakes rule is absolute: even simple changes in critical code use a Titan. A one-line change to authentication logic that looks straightforward to a Sprinter might carry a subtle security implication that only a Titan's deeper reasoning catches.

Phase-Based Allocation: Matching Models to Project Stages

For larger features, optimize model choice by phase of work:

Phase	What You're Doing	Model	Why
Architecture	Schema design, system decomposition, trade-off analysis	Titan	High-leverage decisions. Deep reasoning pays for itself here.
Core Implementation	Feature development, business logic, API building	Workhorse	Bulk of the work. Fast, accurate, economical at scale.
Boilerplate + Docs	Tests for known functions, JSDoc, type interfaces, README	Sprinter	Pattern work. No deep reasoning needed. Save cost and time.
QA + Security Review	Pre-merge security audit, edge case review, final QA	Titan	Critical gate. Catching one security bug here justifies the cost 100×.

The Progressive Escalation Rule

Don't start with a Titan for everything. Use the escalation ladder:

1. Start with Workhorse

First attempt. Covers 80% of tasks.

→

2. Escalate if Stuck

After 2–3 attempts. Bring in Titan.

→

3. Titan for Blockers

Deep problems that need deep reasoning.

This "trust but verify" approach uses Titan-level intelligence only where it's genuinely needed — which is roughly 20% of tasks. The 80% that Workhorse handles correctly? You've just cut your AI costs significantly without any quality loss.

Next in AI Workflow

Part 10 — Team AI

Individual AI workflows don't scale. How to build a shared standard across your entire engineering team — so everyone gets the benefit, not just the developers who figured it out themselves.