Route cheap models for simple tasks — save 60% without losing quality

Claude Haiku 4.5 just changed the agent economics game. It's 4x faster than Sonnet and costs 90% less, but here's what nobody's talking about: most of your agent's work doesn't need Sonnet's power.

I've been running cost routing for three months. My agent bills dropped 60% while quality stayed the same. Here's the pattern that works:

Route by task complexity, not by default model. Let Haiku handle the easy stuff, escalate to Sonnet when it matters.

The key is building a routing layer that makes decisions before the expensive model even sees the request. Here's my routing logic:

def route_model(task_type, context_size, requires_reasoning):
    # Route to Haiku for simple tasks
    if task_type in ['format', 'extract', 'summarize', 'classify']:
        if context_size < 10000 and not requires_reasoning:
            return 'haiku-4.5'
    
    # Route to Sonnet for complex work
    if task_type in ['analyze', 'plan', 'debug', 'create']:
        return 'sonnet-3.5'
    
    # Default to Haiku with escalation
    return 'haiku-4.5-with-fallback'

But routing isn't just about the model — it's about task design. I restructured my agent workflows to create more Haiku-friendly tasks:

• Break complex requests into simple steps — instead of "analyze this code and suggest improvements," I do "extract functions" (Haiku) then "suggest improvements" (Sonnet)
• Use Haiku for validation — "Does this output match the requirements?" costs pennies and catches 80% of issues
• Route status updates to Haiku — "Update the user on progress" doesn't need Sonnet's reasoning power

The biggest win: Haiku with escalation. When Haiku hits something too complex, it calls Sonnet automatically. I track escalation rates — if a task type escalates more than 30%, I route it directly to Sonnet.

# Example escalation prompt for Haiku
"If this task requires complex reasoning, multi-step planning, 
or creative problem-solving, respond with 'ESCALATE: [reason]' 
instead of attempting the task."

Real numbers from my setup: Email processing went from $0.15 per email (Sonnet) to $0.03 per email (Haiku). Code formatting dropped from $0.08 per file to $0.01. The only tasks that stayed on Sonnet: architecture decisions, complex debugging, and creative writing.

Warning: Don't route blindly by cost. I tried routing everything to Haiku first — it created more work fixing bad outputs than I saved on API costs.

The pattern that works: Design for Haiku, escalate to Sonnet. Your agent gets faster, cheaper, and more predictable. Most tasks don't need the expensive model — they just need the right task breakdown.

Start with your highest-volume, lowest-complexity tasks. Route those to Haiku and measure the quality difference. You'll be surprised how little you lose and how much you save.