Use Smaller AI Models for Classification and Routing Tasks

Not every AI task needs the most powerful model. Use cheaper, faster models like gpt-4o-mini or claude-3-5-haiku for classification, routing, extraction, and short summaries, and reserve full-size models for long-form generation and complex reasoning. In n8n, set the model per node so simple decisions cost a fraction of a cent.

When should you use a smaller AI model?

Not every AI task needs the most powerful (and expensive) model. Use cheaper, faster models for classification, routing, extraction, and yes/no decisions. Reserve the full-sized models for tasks that genuinely need them -- long-form generation, complex reasoning, or nuanced analysis.

Real-world example: A support ticket workflow classifies incoming emails into categories (billing, technical, feature request, spam), then routes them to the right team. Classification is a simple task that a small model handles perfectly.

Task-to-model mapping:

Classification/routing:
  Model: gpt-4o-mini or claude-3-5-haiku
  Cost: ~$0.0001 per classification
  Example: "Is this email about billing, technical support, or a feature request?"

Data extraction:
  Model: gpt-4o-mini or claude-3-5-haiku
  Cost: ~$0.0003 per extraction
  Example: "Extract the customer name, order ID, and issue from this email."

Summarization (short):
  Model: gpt-4o-mini or claude-3-5-haiku
  Cost: ~$0.0005 per summary
  Example: "Summarize this support ticket in one sentence."

Long-form generation:
  Model: gpt-4o or claude-sonnet-4
  Cost: ~$0.01-0.05 per generation
  Example: "Write a detailed response to this customer's technical issue."

Complex reasoning:
  Model: gpt-4o or claude-sonnet-4
  Cost: ~$0.02-0.10 per query
  Example: "Analyze this contract clause and identify potential risks."

How do you set the model per task in n8n?

In n8n, configure the OpenAI or Anthropic node to use the appropriate model per task:

// Classification node (cheap model)
{
  "resource": "chat",
  "model": "gpt-4o-mini",
  "messages": [
    {
      "role": "system",
      "content": "Classify the following support email into exactly one category: billing, technical, feature_request, spam. Respond with only the category name."
    },
    {
      "role": "user",
      "content": "={{ $json.emailBody }}"
    }
  ],
  "temperature": 0,
  "maxTokens": 10
}

The generation node uses the powerful model, and only when it is actually needed:

// Generation node (powerful model, only called when needed)
{
  "resource": "chat",
  "model": "gpt-4o",
  "messages": [
    {
      "role": "system",
      "content": "Write a helpful, empathetic response to this customer support ticket. Include specific steps to resolve their issue."
    },
    {
      "role": "user",
      "content": "={{ $json.emailBody }}"
    }
  ],
  "temperature": 0.7,
  "maxTokens": 500
}

How much does smart routing save?

Cost comparison for 1000 support tickets/day:
  All gpt-4o:      1000 x $0.03 = $30/day = $900/month
  Smart routing:    1000 x $0.0001 (classify) + 200 x $0.03 (generate) = $6.10/day = $183/month
  Savings: ~80%

The key insight: most tickets only need classification and a template response. Only the complex ones need a powerful model to generate a custom reply.

Related: Use Structured Output (JSON Mode) for Parseable Responses · Configure Payload Size and Binary Data Mode for Large Files