kienbui1995/cost-aware-routing

Cost-Aware Model Routing

Overview

Not every task needs the most expensive model. Route tasks to the right model based on complexity, saving cost without sacrificing quality.

The Routing Table

| Model | Agent | Cost | Best For | |-------|-------|------|----------| | Opus | architect | $$$$$ | Complex architecture, novel design, ambiguous requirements | | Sonnet | debugger, ui-designer, (default) | $$ | Implementation, debugging, most coding tasks | | Haiku | reviewer | $ | Code review, verification, simple lookups, formatting |

Cost ratios (approximate):

• Opus : Sonnet : Haiku ≈ 25 : 5 : 1
• A Haiku review costs ~1/60th of an Opus review

Decision Framework

WHEN receiving a task:

1. Can Haiku handle this? (review, verify, format, lookup)
   → YES: Use reviewer agent or --model haiku
   
2. Is this standard implementation/debugging?
   → YES: Use default Sonnet (or debugger/ui-designer agent)
   
3. Does this require deep reasoning about novel problems?
   → YES: Use architect agent (Opus)

Task → Model Mapping

Use Haiku (reviewer agent) for:

• Code review after implementation
• Verification checks (tests pass? linter clean?)
• Simple refactoring (rename, reformat)
• Documentation review
• Checking git status/diff
• Quick lookups in codebase

Use Sonnet (default) for:

• Feature implementation
• Bug fixing and debugging
• Writing tests
• Database migrations
• API development
• UI component building
• Most day-to-day coding

Use Opus (architect agent) for:

• System architecture design
• Complex brainstorming with many trade-offs
• Ambiguous requirements that need deep analysis
• Cross-system integration planning
• Performance optimization strategy
• Security architecture review

Anti-Patterns

| Pattern | Problem | Fix | |---------|---------|-----| | Using Opus for everything | 25x cost for same result | Default to Sonnet | | Using Opus for code review | Overkill, slow | Use Haiku reviewer | | Using Haiku for implementation | Too many mistakes | Use Sonnet | | Not using agents at all | Missing cost savings | Dispatch to right agent |

CLI Shortcuts

# Quick review with Haiku
claude -p "review the last commit" --model haiku

# Debug with Sonnet (default)
claude -p "fix the failing test in auth.py"

# Architecture with Opus
claude -p "design the caching layer" --model opus

# Budget cap
claude --max-budget-usd 0.50 -p "implement the login form"

Measuring Impact

Track your usage in Claude Code settings or Vertex AI console:

• Before: ~98% Opus = high cost
• After: ~10% Opus, ~70% Sonnet, ~20% Haiku = ~75% cost reduction