Pick Model - Extended Reference

Model Characteristics

Haiku 4.5

• Speed: Fastest (~2-3x faster than Sonnet)
• Cost: Lowest (~10x cheaper than Opus)
• Context: 200K tokens
• Best for: Deterministic, pattern-based, low-reasoning tasks
• Limitations: Struggles with ambiguity, multi-step reasoning, creative tasks

Sonnet 4.5

• Speed: Medium (baseline)
• Cost: Medium (~5x cheaper than Opus)
• Context: 200K tokens
• Best for: Balanced reasoning, creative work, most coding tasks
• Limitations: Multi-file refactoring, highly nuanced reasoning

Opus 4.6

• Speed: Slowest (~1.5-2x slower than Sonnet)
• Cost: Highest (premium tier)
• Context: 200K tokens
• Best for: Complex reasoning, architectural decisions, high-stakes work
• Limitations: Overkill for simple tasks, slower iteration

Extended Decision Matrix

By File Type

By Domain

By Interaction Pattern

Cost/Latency Tradeoffs

When to optimize for speed (choose lower tier)

• Rapid prototyping, quick iteration
• Low-stakes exploratory work
• User waiting synchronously
• Batch processing many simple tasks

When to optimize for quality (choose higher tier)

• Production deployments
• Security-critical code
• User-facing content (brand reputation)
• Complex architectural decisions
• Tasks where rework is expensive

Cost Examples (Approximate)

• 100K input tokens + 10K output:
  • Haiku: ~$0.10
  • Sonnet: ~$0.30
  • Opus: ~$1.50

Edge Cases & Hybrid Tasks

Escalation Scenarios

Start Haiku → Upgrade Sonnet if:

• Output lacks coherence
• Task requires reasoning not obvious from pattern
• User requests "explain why" or "consider alternatives"

Start Sonnet → Upgrade Opus if:

• Multi-system dependencies emerge
• Ambiguity requires nuanced judgment
• Initial approach fails, root cause unclear
• Architectural implications surface

Hybrid Approaches

Sequential (pipeline):

1. Haiku: Extract data from logs
2. Sonnet: Analyze patterns, generate report

Parallel (fan-out):

1. Haiku: Format 10 files in parallel
2. Sonnet: Review aggregated changes

Iterative (feedback loop):

1. Sonnet: Draft implementation plan
2. User: Feedback
3. Opus: Refine with architectural considerations

Domain-Specific Guidelines

Web Development

• Component styling, prop changes → Sonnet
• Component library design → Opus
• API endpoint (single) → Sonnet
• API architecture (REST vs GraphQL) → Opus

Data Science

• Data cleaning, feature engineering → Haiku/Sonnet
• Model selection, experiment design → Opus
• Jupyter notebook fixes → Sonnet
• Pipeline architecture → Opus

Infrastructure

• Terraform syntax fix → Haiku
• Resource provisioning → Sonnet
• Multi-region HA design → Opus

Documentation

• API reference generation → Haiku
• Tutorial writing → Sonnet
• Architecture Decision Records → Opus

Common Mistakes

Over-escalation

• Using Opus for typo fixes, simple formatting
• Cost: 10-15x more expensive
• Fix: Trust Haiku for deterministic tasks

Under-estimation

• Using Haiku for "simple" refactors that touch 5+ files
• Risk: Poor code quality, missed edge cases
• Fix: Apply complexity escalators

Ignoring context

• Choosing model without considering stakes, ambiguity, scope
• Fix: Use decision matrix + escalators

False economy

• Choosing Haiku for production-critical work to save $1
• Risk: Outages, security issues, rework costs >> savings
• Fix: Escalate +1 tier for high stakes

When to Override Recommendation

User knows best when:

• Specific model preferences based on past experience
• Budget constraints require cost optimization
• Time constraints require speed optimization
• Iterative work (start lower, escalate if needed)

Always respect explicit user model selection.

Quick Reference: Signal Words

Haiku signals

• "quick", "simple", "just", "only", "extract", "format", "rename", "fix typo"

Sonnet signals

• "write", "create", "explain", "review", "analyze", "debug", "single file"

Opus signals

• "design", "architect", "complex", "multiple files", "refactor", "migration", "strategy", "nuanced"

Model Selection Confidence

Output confidence level based on task clarity:

🔴 **Opus** — [reason]
💰 Cost: highest | ⚡ Speed: slowest
🎯 Confidence: High (clear architectural scope)

💡 Consider Sonnet if scope reduces to 2-3 files after exploration

Confidence indicators:

• High: Task signals clearly match one tier
• Medium: Borderline between two tiers
• Low: Insufficient information, recommend exploration </content>