Lesson 8: AI Model Selection
Time: 25 minutes Goal: Optimize cost and quality by choosing the right model
Claude Models
| Model | Best For | Cost | Speed |
|---|---|---|---|
| Opus 4.5 | Architecture, complex analysis | $$$ | Slower |
| Sonnet 4.5 | Daily coding, features, tests | $$ | Fast |
| Haiku 3.5 | Quick lookups, simple edits | $ | Fastest |
The most expensive model isn't always best. Using Opus for a typo is like a sledgehammer for a thumbtack.
When to Use Each Model
Opus 4.5
✅ Architecture decisions
✅ Complex multi-file refactoring
✅ System design
✅ Security reviews
✅ Comprehensive specs
Example commands:
/specweave:increment "..." # Uses Opus for planning
/specweave:qa 0001 # Uses Opus for analysis
Sonnet 4.5
✅ Feature implementation
✅ Writing tests
✅ Bug fixes
✅ Code reviews
✅ Daily development
Example commands:
/specweave:do # Uses Sonnet for execution
Haiku 3.5
✅ Quick lookups
✅ Simple file edits
✅ Bulk find-and-replace
✅ Syntax checks
Example:
"What's the export name in utils/constants.ts?"
"Change button color from blue to green"
SpecWeave Default Selection
SpecWeave picks automatically:
/specweave:increment → Opus (complex planning)
/specweave:do → Sonnet (balanced execution)
/specweave:qa → Opus (deep analysis)
/specweave:sync-progress → Haiku (fast sync)
Cost Optimization
Strategy 1: Progressive Escalation
Start cheap, escalate as needed:
Step 1: Haiku exploration
"What files handle authentication?"
Cost: ~$0.001
Step 2: Sonnet analysis
"Explain the auth flow"
Cost: ~$0.02
Step 3: Opus architecture (if needed)
"Redesign for multi-tenant"
Cost: ~$0.15
Strategy 2: Batch Similar Tasks
❌ Expensive: 10 separate Opus calls
"Fix typo in file1.ts"
"Fix typo in file2.ts"
...
✅ Cheap: 1 Haiku bulk operation
"Fix all typos: file1.ts, file2.ts, ..."
Strategy 3: Explore First
# Cheap exploration with Haiku
/specweave:do --explore-only
# Then implement with Sonnet
/specweave:do --continue
Real-World Example
Task: "Add payment processing"
| Phase | Model | Cost | Action |
|---|---|---|---|
| Research | Haiku | $0.005 | Find existing code |
| Planning | Opus | $0.20 | Create increment |
| Implementation | Sonnet | $1.20 | Build 8 tasks |
| Total | $1.40 | vs $3.00 all-Opus |
Configuration
Default Models
In .specweave/config.json:
{
"ai": {
"defaultModel": "sonnet",
"planningModel": "opus",
"explorationModel": "haiku"
}
}
Per-Increment Override
In spec.md frontmatter:
---
increment: 0001-complex-refactor
ai:
model: opus # Use Opus throughout
---
Quick Reference
┌─────────────────────────────────────────────────────────┐
│ WHICH MODEL? │
├─────────────────────────────────────────────────────────┤
│ │
│ "I need to..." │
│ │
│ Design / Architect ────────────► OPUS 4.5 │
│ Complex refactor │
│ Security review │
│ │
│ Write features ────────────► SONNET 4.5 │
│ Fix bugs │
│ Write tests │
│ │
│ Quick lookup ────────────► HAIKU 3.5 │
│ Simple edit │
│ Bulk operation │
│ │
└─────────────────────────────────────────────────────────┘
Glossary Terms Used
Key Takeaways
- Start cheap (Haiku), escalate as needed
- Use Opus for planning, Sonnet for execution
- Batch similar operations
- Cache exploration results
What's Next?
Learn how to fix common issues.
:next → Lesson 9: Troubleshooting