Model Selection Decision Framework

Choosing the right model is not a technical decision — it is an economic one. This lesson gives you a decision framework that optimizes for total cost of ownership, not just API cost.

The True Cost Equation

Most people only consider API cost. The real equation is:

Total Cost = API Cost + Human Review Cost + Error Cost + Opportunity Cost

• API Cost: Token pricing (Opus is ~1.7× Sonnet, ~20× Haiku)
• Human Review Cost: Time spent checking and fixing AI output
• Error Cost: Production bugs, security incidents, data corruption
• Opportunity Cost: Slower development velocity from using a weaker model

For critical tasks, Opus 4.6 is often the cheapest option because it minimizes human review and error costs.

Decision Matrix

graph TD
    A[New Task] --> B{Correctness<br/>Critical?}
    B -->|Yes| C{Budget<br/>Constrained?}
    B -->|No| D{Volume<br/>>100/day?}
    C -->|No| E[Opus 4.6]
    C -->|Yes| F[Opus 4.6 with<br/>cost controls]
    D -->|Yes| G[Haiku]
    D -->|No| H{Complexity?}
    H -->|Simple| G
    H -->|Medium| I[Sonnet 4.5]
    H -->|Complex| E

Task-to-Model Mapping

Task Category	Model	Reasoning
Security audit	Opus 4.6	Error cost is catastrophic
Architecture design	Opus 4.6	Requires deep multi-step reasoning
Code review (critical)	Opus 4.6	Must catch subtle bugs
Code review (routine)	Sonnet 4.5	Good enough for style and basic issues
Bug fixing (complex)	Opus 4.6	Better at root cause analysis
Bug fixing (simple)	Sonnet 4.5	Cost-effective for straightforward fixes
Documentation	Sonnet 4.5	Quality sufficient, lower cost
Code completion	Sonnet 4.5	Speed matters more than depth
Log parsing	Haiku	High volume, pattern matching
Data extraction	Haiku	Structured, repetitive tasks
Customer support triage	Haiku	Simple classification

Building an Automated Model Router

from dataclasses import dataclass
from enum import Enum

class ModelTier(Enum):
    OPUS = "claude-opus-4-6-20260205"
    SONNET = "claude-sonnet-4-5-20241022"
    HAIKU = "claude-haiku-3-5-20241022"

@dataclass
class TaskProfile:
    correctness_critical: bool
    complexity: str  # "low", "medium", "high"
    volume: str      # "single", "batch", "stream"
    domain: str      # "security", "coding", "content", "data"

def route_to_model(task: TaskProfile) -> ModelTier:
    """Route tasks to the appropriate model tier."""
    # Security and architecture always get Opus
    if task.domain in ("security", "architecture"):
        return ModelTier.OPUS

    # Critical correctness gets Opus
    if task.correctness_critical and task.complexity != "low":
        return ModelTier.OPUS

    # High volume simple tasks get Haiku
    if task.volume == "stream" and task.complexity == "low":
        return ModelTier.HAIKU

    # Everything else gets Sonnet
    if task.complexity == "high":
        return ModelTier.OPUS

    return ModelTier.SONNET

Cost Projection Example

A team of 5 engineers using Claude daily:

Scenario	Opus Only	Smart Routing	Savings
500 requests/day	$125/day	$45/day	64%
Monthly cost	$3,750	$1,350	$2,400/month

Smart routing sends ~20% of tasks to Opus, ~50% to Sonnet, ~30% to Haiku — without sacrificing quality where it matters.

In the next lesson, we cover the breaking API changes when migrating from Opus 4.5.