LLM Routing: How to Cut 40% of Your Agent Bill in 10 Lines

The cost anti-pattern\n\nMost production agent loops I see call one model for everything: claude-opus-4-6 or gpt-5. The reasoning: *Opus is smarter so the output quality is higher and retries cost less.*\n\nThat's true for complex reasoning. It's catastrophically wrong for the 70% of prompts that are trivial — summarization, classification, routing, extraction, one-shot transforms. You're paying premium prices for tasks that Haiku or Gemini Flash would nail perfectly.\n\n## The simplest possible router\n\nA 10-line prompt classifier gives you 80% of the savings. Route short + simple to Haiku, medium to Sonnet, complex to Opus. Replace with GPT / Gemini — same structure.\n\nTypical savings: 35–50% vs always-Opus.\n\n## Go deeper gradually\n\nBetter signals to add over time: previous turn failed → upgrade. Token budget remaining → downgrade. User tier → quality floor. Latency SLA → fastest above quality bar.\n\nAll of this is covered by lazymac LLM Router — free API that returns a model pick with a reason.\n\n``bash\ncurl -X POST https://api.lazy-mac.com/llm-router/api/v1/route -d '{\"prompt\":\"Summarize this\",\"min_quality\":80}'\n``\n\n## Measuring the savings\n\nYou can't improve what you can't measure. The AI Spend Tracker logs every call with model + tokens + cost, so you can prove the 40% number on your own workload.\n\nFree: 100 logs/day. Pro $29/mo unlimited + budgets + alerts + forecast.\n\n## The Pareto\n\n1. Route short + simple prompts to Haiku / Gemini Flash\n2. Cap multi-turn loops at N iterations with cost-aware fallback\n\nEverything past that is diminishing returns.

Support the work: MCP Pro $29/mo · MCPWatch Pro Report $49 · more posts