model-selection-guide

Model Selection Guide

When to Use

• Unsure which model to use for a specific task
• Want to reduce cost without sacrificing quality
• Designing a multi-phase workflow with model assignments per phase
• @workflow-model-advisor uses this as its decision framework

Core Jobs

1. Decision Tree

Step 1: Identify task type

Is this primarily reading + pattern matching + structured output?
  → Haiku (fast, cheap, accurate for structured tasks)
  Examples: code review, lint check, classify intent, format output,
            summarize doc, extract fields, check compliance

Is this implementation, debugging, or moderate reasoning?
  → Sonnet (balanced quality + cost)
  Examples: write code, fix bug, explain concept, write tests,
            analyze trade-offs (simple), build feature

Is this deep architecture, novel problem, or complex multi-step reasoning?
  → Opus (highest capability, use sparingly)
  Examples: system design, architecture decisions, evaluate complex trade-offs,
            research synthesis, "should we X or Y?" with many unknowns

Step 2: Apply cost heuristic

Model	Relative cost	Speed	Best for
Haiku	$ (1x)	Fastest	Review, classify, format, extract
Sonnet	$$ (5x)	Fast	Implement, debug, explain, write
Opus	$$$$ (15x)	Slower	Design, architecture, deep analysis

Step 3: Check override conditions

Use Opus even for "simple" tasks if:

• Task has irreversible consequences (production changes, data migrations)
• Previous Sonnet attempt produced insufficient output

Use Haiku even for "complex" tasks if:

• Output is structured/templated (Haiku excels at templates)
• Task is repetitive across many items (batch processing)
• Speed is critical and quality threshold is moderate

2. Per-Workflow Model Assignments

Workflow	Phase	Agent	Model	Why
feature	plan	@architect	Opus	Architecture decisions need deep reasoning
feature	implement	subagent	Sonnet	Code generation, balanced
feature	test	subagent	Sonnet	Test writing needs understanding
feature	review	@reviewer	Haiku	Pattern matching, structured output
bugfix	reproduce	subagent	Sonnet	Needs to understand codebase
bugfix	diagnose	@debugger	Sonnet	Systematic debugging
bugfix	fix	subagent	Sonnet	Code fix, minimal
bugfix	verify	@reviewer	Haiku	Check fix completeness
refactor	analyze	@architect	Opus	Understanding complex structure
refactor	plan	@architect	Opus	Design target state
refactor	implement	subagent	Sonnet	Mechanical refactoring
refactor	review	@reviewer	Haiku	Pattern check
research	gather	subagent	Sonnet	Information gathering
research	synthesize	@architect	Opus	Complex reasoning over options
research	document	@technical-writer	Haiku	Structured doc output
incident	triage	@debugger	Sonnet	Fast diagnosis
incident	hotfix	subagent	Sonnet	Minimal fix
incident	postmortem	@technical-writer	Haiku	Structured template

3. Cascade Pattern

When uncertain, start cheaper and escalate only if needed:

1. Try Haiku → evaluate output quality
2. If insufficient → escalate to Sonnet → evaluate
3. If still insufficient → escalate to Opus
4. Never auto-escalate without notifying user of cost increase

4. Cost Estimates Per Task

Task type	Model	Estimated cost
Code review (500 lines)	Haiku	~$0.01
Bug fix (small)	Sonnet	~$0.05-0.15
Feature implementation	Sonnet	~$0.10-0.50
Architecture design	Opus	~$0.50-2.00
Research synthesis	Opus	~$0.50-1.50

Cost reminder: Opus = 15x Haiku. Code review in Opus costs 15x more with same quality output.

Key Concepts

• Haiku — fast, cheap, accurate for structured/pattern tasks; NOT for deep reasoning
• Sonnet — balanced for most implementation; default for unknown tasks
• Opus — most capable; use only when task genuinely requires deep reasoning
• Cascade — start cheap, escalate if output quality insufficient
• Cost multiplier — Opus = 15x Haiku; always ask "does this task need Opus?"

Checklist

• Task type identified before selecting model?
• Haiku considered first for review/classification/formatting tasks?
• Opus reserved for genuine architecture/deep-reasoning tasks?
• User notified if escalating from Haiku to Opus (15x cost increase)?
• Per-phase model assignments made for multi-phase workflows?

Key Outputs

• Model recommendation with reasoning + cost estimate
• Cascade suggestion if task type is ambiguous

Output Format

• 🔴 Critical — using Opus for code review or formatting (massive overspend), using Haiku for architecture decisions (under-powered)
• 🟡 Warning — defaulting to Sonnet for everything without considering Haiku for structured tasks
• 🟢 Suggestion — use cascade pattern for uncertain tasks, assign Haiku to all review phases

Anti-Patterns

• Using Opus as the safe default (expensive, often unnecessary)
• Never using Haiku (leaving 60x savings on the table for review tasks)
• Using model complexity as proxy for task importance
• Switching models mid-task without reason (inconsistent output)

Integration

• workflow-templates — model assignments per phase come from this skill
• @workflow-model-advisor uses this as its decision framework
• @workflow-orchestrator reads model assignments when dispatching phase subagents