Smart LLM Routing: Using OpenRouter with Mithril for Cost-Optimized AI

Not every agent task needs the most expensive LLM. A simple classification might use Haiku, while a complex analysis needs Opus. OpenRouter + Mithril lets your agent make this choice dynamically — and pay per-call for only what it uses.

Why LLM Routing Matters

LLM costs vary 100x:

An agent that uses Opus for everything costs 50x more than one that routes intelligently.

OpenRouter: The Universal LLM API

OpenRouter provides a single API endpoint that routes to any LLM. Instead of integrating with Anthropic, OpenAI, Google, and Meta separately, you call OpenRouter and specify the model.

With x402, your agent doesn't even need an OpenRouter API key. It just pays per-call.

Routing Strategy

function selectModel(task: string): string {
  // Simple classification/extraction
  if (task === "classify" || task === "extract")
    return "anthropic/claude-haiku-4-5"

  // Standard analysis/writing
  if (task === "analyze" || task === "write")
    return "anthropic/claude-sonnet-4-6"

  // Complex reasoning/planning
  if (task === "plan" || task === "reason")
    return "anthropic/claude-opus-4-6"

  // Default to cost-effective
  return "anthropic/claude-sonnet-4-6"
}

Implementation

async function callLLM(task: string, prompt: string) {
  const model = selectModel(task)

  const result = await m.pay({
    url: "https://openrouter.ai/api/v1/chat/completions",
    method: "POST",
    body: JSON.stringify({
      model,
      messages: [{ role: "user", content: prompt }],
    }),
  })

  return result.data.choices[0].message.content
}

Cost Impact

A research agent that makes 100 LLM calls/day:

|----------|-----------|

Smart routing saves 60-92% on LLM costs by matching model capability to task complexity.

Advanced: Dynamic Model Selection

Let the agent itself decide which model to use:

async function smartCall(prompt: string) {
  // Use a cheap model to classify the task
  const classification = await callLLM(
    "classify",
    `Classify this task as simple/medium/complex: ${prompt}`
  )

  const model = classification.includes("complex")
    ? "anthropic/claude-opus-4-6"
    : classification.includes("medium")
    ? "anthropic/claude-sonnet-4-6"
    : "anthropic/claude-haiku-4-5"

  return callLLM("analyze", prompt)
}

The classification call costs ~$0.001 and can save $0.05+ on every subsequent call.

Monitoring

Track model usage in the Mithril dashboard: