The 80/20 Rule of AI Model Selection (Why You're Overpaying)

80% of your AI API calls probably don't need a frontier model.

I spent a month tracking every API call in my workflow. The breakdown was almost comically predictable:

• 40% simple tasks: formatting, imports, typo fixes, boilerplate generation
• 35% standard tasks: refactoring, code review, test writing, documentation
• 20% complex tasks: architecture decisions, multi-file debugging, system design
• 5% genuinely hard tasks: novel algorithms, cross-system integration, performance optimization

Only that bottom 25% actually benefits from Opus-class reasoning. The top 75% runs identically on cheaper models.

the cost math

Task Type	Best Model	Cost per M tokens	% of calls
Simple	Haiku	$0.25 / $1.25	40%
Standard	Sonnet	$3 / $15	35%
Complex	Opus	$5 / $25	20%
Hard	Opus (max effort)	$5 / $25	5%

Weighted average cost: ~$2.60 / $12.25 per M tokens
All-Opus cost: $5 / $25 per M tokens

Savings: ~48% on input, ~51% on output.

And that's conservative. If you're aggressive about routing simple stuff to Haiku, savings can hit 60-70%.

how to actually do this

Three approaches, from simplest to most automated:

1. Manual routing (free, 5 minutes)

Just be intentional about which model you call. Before every API request, ask: does this need Opus?

def pick_model(task_description):
    simple_keywords = ["fix", "format", "import", "typo", "rename"]
    if any(k in task_description.lower() for k in simple_keywords):
        return "claude-haiku-4-5"
    return "claude-sonnet-4-6"  # default to Sonnet, not Opus

2. Keyword-based routing (free, 30 minutes)

Build a classifier based on prompt characteristics. Long prompts with multiple files = Opus. Short prompts with clear intent = Haiku.

3. Automated routing (various tools)

Several tools now handle this automatically. They classify prompt complexity and route to the cheapest model that can handle it. The category is called "LLM routing" or "model routing" if you want to explore options.

the uncomfortable truth

Most developers default to the most expensive model because switching feels risky. But the risk is actually backwards: you're guaranteed to overpay with a single-model strategy. With routing, the worst case is Opus handles everything (same as before). The best case is 50-70% savings.

The 80/20 rule applies perfectly here. 80% of your AI spend is going to the 20% of calls that don't need it.

---

Part of my series on AI cost optimization. See also: Cut Your Claude Bill 60%, Opus 4.7 Token Analysis, Hidden Effort Level Setting