We cut our agent costs 60% by removing model loyalty, not adding efficiency

Last month we were spending $847/month on Claude Sonnet for everything. Code reviews, email drafts, data analysis, customer support—one model, premium pricing, because "Sonnet is the best."

Then GPT-6 Astra dropped. Claude Sonnet 4 followed two weeks later. Gemini Flash got a massive update. Instead of picking a winner, we built a router.

Now we spend $312/month and get better results.

The trick isn't loyalty—it's routing by task complexity, not model preference.

The routing pattern that cuts costs 60%: Match task complexity to model capability, not brand loyalty.

Here's our routing config:

SIMPLE_TASKS = [
    "email_drafts", "basic_summaries", "formatting", 
    "simple_rewrites", "status_updates"
]
route_to: gemini-flash-2.0
cost: $0.15/1M tokens

COMPLEX_TASKS = [
    "architecture_decisions", "code_reviews", 
    "strategic_analysis", "debugging"
]
route_to: claude-sonnet-4
cost: $3.00/1M tokens

CREATIVE_TASKS = [
    "marketing_copy", "product_naming", 
    "user_stories", "brainstorming"
]
route_to: gpt-6-astra
cost: $2.40/1M tokens

The math is brutal when you track it. 70% of our agent work was simple tasks getting routed to a $3/1M token model. Flash handles email drafts and formatting for $0.15/1M tokens—20x cheaper with identical quality.

But routing by price alone breaks everything. We tried sending code reviews to Flash first. Disaster. It missed edge cases that Sonnet caught immediately. The $2.85 we saved per review cost us 3 hours of debugging.

Task complexity beats model rankings. Sonnet 4 is "better" than Astra on benchmarks, but Astra writes marketing copy that actually converts. Flash is "worse" than both, but it formats data perfectly and costs almost nothing.

Our router now tracks completion success, not just token costs:

task_routing = {
    "completion_rate": 0.94,  # Flash for simple tasks
    "revision_count": 1.2,    # Average revisions needed
    "cost_per_completion": 0.08,  # Total cost including retries
    "fallback_trigger": "complexity_score > 0.6"  # Auto-escalate
}

The fallback pattern matters more than the routing. When Flash hits something complex, it escalates to Sonnet automatically. When Sonnet gets stuck on creative work, it hands off to Astra. No model loyalty, just task-matched intelligence.

Three months in, our agents handle 40% more work for 63% less cost. The router saved us $1,608 this quarter.

Model wars are marketing. Task routing is math.