Retour aux articles
5 MIN READ

LLM Routing: Choosing the Right Model for Each Task

By Learnia Team

LLM Routing: Choosing the Right Model for Each Task

This article is written in English. Our training modules are available in French.

Should every question go to GPT-4? That's like using a sports car to fetch groceries—overkill for simple tasks and unnecessarily expensive. LLM routing matches questions to the right model, optimizing cost and speed without sacrificing quality.


What Is LLM Routing?

LLM routing is the practice of directing different queries to different AI models based on task requirements.

The Basic Concept

User query → Router → Appropriate model

"What's 2+2?" → Fast, cheap model (GPT-3.5)
"Analyze this legal contract" → Powerful model (GPT-4)
"Generate a poem" → Creative model (Claude)

Why Routing Matters

The Cost Reality

| Model | Input Cost (per 1M tokens) | Quality | |-------|---------------------------|---------| | GPT-3.5 Turbo | $0.50 | Good | | GPT-4 Turbo | $10.00 | Excellent | | GPT-4o | $2.50 | Very Good | | Claude 3 Haiku | $0.25 | Good | | Claude 3 Opus | $15.00 | Excellent |

20-60× price difference between models.

The Math

Without routing:
1000 queries/day × GPT-4 ($0.01/query) = $10/day = $300/month

With routing (70% simple, 30% complex):
700 queries × GPT-3.5 ($0.0005) = $0.35
300 queries × GPT-4 ($0.01) = $3.00
Total: $3.35/day = $100/month

Savings: 67% cost reduction

Routing Strategies

1. Task-Based Routing

Route based on what the user is asking:

Classification/extraction → Small model
Creative writing → Medium model
Complex reasoning → Large model
Code generation → Specialized model

2. Complexity-Based Routing

Estimate query difficulty:

Simple: "What's the weather?"
→ Fast model

Medium: "Summarize this article"
→ Balanced model

Complex: "Compare these three legal arguments"
→ Powerful model

3. Cascade Routing

Try smaller model first, escalate if needed:

Step 1: Send to GPT-3.5
Step 2: Check confidence/quality
Step 3: If uncertain → re-send to GPT-4

4. Intent-Based Routing

Classify intent, then route:

Intent: customer_support → Support-tuned model
Intent: code_help → Code-specialized model
Intent: creative → Creative model
Intent: analysis → Reasoning model

What Makes a Query "Complex"?

Signals of Complexity

✓ Multi-step reasoning required
✓ Domain expertise needed
✓ Long context to process
✓ Nuanced judgment required
✓ High-stakes outcome

Signals of Simplicity

✓ Single fact lookup
✓ Simple format conversion
✓ Short, clear instruction
✓ Low-stakes outcome
✓ Well-defined output

Real-World Routing Examples

Customer Support Bot

"What are your hours?"
→ Route to FAQ lookup + small model
   Cost: $0.0001 | Latency: 200ms

"I'm having a complex billing dispute about..."
→ Route to support-specialized model + human escalation flag
   Cost: $0.005 | Latency: 1s

Code Assistant

"Add a comment to this line"
→ Small, fast model
   Cost: $0.0002

"Refactor this 500-line function for performance"
→ Large model with long context
   Cost: $0.02

Research Assistant

"When was the Eiffel Tower built?"
→ Small model (factual recall)

"Compare the economic impacts of three trade policies"
→ Large model (analysis + reasoning)

Cascade Pattern Deep Dive

The cascade approach is particularly powerful:

┌─────────────────┐
│ User Query      │
└────────┬────────┘
         ▼
┌─────────────────┐
│ Tier 1: Small   │──── Confident? ──▶ Return answer
│ (GPT-3.5/Haiku) │         │
└────────┬────────┘         │ No
         ▼                  ▼
┌─────────────────┐
│ Tier 2: Medium  │──── Confident? ──▶ Return answer
│ (GPT-4o/Sonnet) │         │
└────────┬────────┘         │ No
         ▼                  ▼
┌─────────────────┐
│ Tier 3: Large   │────────────────▶ Return answer
│ (GPT-4/Opus)    │
└─────────────────┘

Advantages of Cascading

✓ Most queries resolved by cheap model
✓ Complex queries still get best model
✓ Natural quality/cost optimization
✓ Built-in escalation path

Implementing a Simple Router

Conceptual Approach

1. Analyze incoming query
   - Length
   - Keywords (e.g., "analyze", "compare", "simple")
   - Domain detection

2. Assign complexity score (0-10)
   - 0-3: Simple → Small model
   - 4-6: Medium → Medium model  
   - 7-10: Complex → Large model

3. Route to selected model

4. (Optional) Evaluate response quality
   - If low quality, retry with larger model

Routing Signals

Simple query indicators:
- Short (< 20 words)
- Contains "what is", "define", "when"
- Single question

Complex query indicators:
- Long (> 100 words)
- Contains "analyze", "compare", "evaluate"
- Multiple sub-questions
- Technical jargon
- Attached documents

Common Routing Mistakes

1. Over-Routing to Expensive Models

❌ Send everything to GPT-4 "just in case"
✅ Trust smaller models for simple tasks

2. Under-Routing Complex Tasks

❌ Always use the cheapest model
✅ Accept higher cost for quality-critical tasks

3. Ignoring Latency

❌ Route based only on cost
✅ Consider: Simple queries need fast responses

4. No Fallback

❌ Single model, no backup
✅ Have escalation path when confidence is low

Key Takeaways

  1. LLM routing matches queries to appropriate models
  2. Can reduce costs by 50-70% without quality loss
  3. Route by task type, complexity, or cascade
  4. Simple queries → cheap/fast models; complex → powerful models
  5. Monitor and iterate—routing rules need tuning

Ready to Build Smart AI Workflows?

This article covered the what and why of LLM routing. But production routing systems require implementation patterns, monitoring, and optimization.

In our Module 4 — Chaining & Routing, you'll learn:

  • Designing multi-model architectures
  • Implementing routing logic
  • Cascade patterns and fallbacks
  • Cost optimization strategies
  • Monitoring and improving routing accuracy

Explore Module 4: Chaining & Routing

GO DEEPER

Module 4 — Chaining & Routing

Build multi-step prompt workflows with conditional logic.