CHAPTER 3.6 · FLEXIBLE INFERENCE-SCALING

Effort Parameter

Explore Claude 4.5's Effort Parameter: Set the reasoning intensity from 1 (Fast) to 10 (Deep) and see in real-time how response time, token costs, and answer quality scale.

Effort Parameter: Claude 4.5's approach to flexible inference — a slider from 1-10 that controls the thinking budget. Simple questions get little compute, complex ones get more.

📖 Learning Context ▼
🎯

Learning Objectives

  • Understand and apply Effort Parameter
  • Optimize cost-quality trade-off
  • Know when which effort level makes sense
🗺️

Context: Where are we?

Step 4/5 Reasoning & Test-Time Compute

Practical application of Test-Time Scaling for developers.

💡

Why it matters

API costs are proportional to thinking tokens. With Effort=3 instead of Effort=10 you save ~70% cost with only ~10% quality loss for simple tasks.

📌

Key Takeaways

  • 1-3: Fast answers, low costs
  • 4-7: Balance for most tasks
  • 8-10: Deep reasoning for complex problems
5
Balanced
Estimated Response Time
15 Seconds
Token Costs
~1,000 Tokens
Answer Quality
⭐⭐⭐⭐
Reasoning Depth
Moderate
📋 Example: Math Problem
Question:
Calculate the number of solutions for:
x³ - 6x² + 11x - 6 = 0
▼ Claude 4.5 with Effort 5
The answer: 3 solutions

(Directly calculated)
Analysis and Solution:

Step 1: Recognize Factorization
I check whether the cubic equation is factorable.

Step 2: Test Rational Roots
Possible rational roots: ±1, ±2, ±3, ±6

Step 3: Substitute and Verify
• x = 1: 1 - 6 + 11 - 6 = 0 ✓
• x = 2: 8 - 24 + 22 - 6 = 0 ✓
• x = 3: 27 - 54 + 33 - 6 = 0 ✓

Step 4: Factorization
(x - 1)(x - 2)(x - 3) = 0

Solution:
The equation has exactly 3 solutions: x = 1, x = 2, x = 3
💡 What changes with the Effort Parameter?

• Effort 1-3: Fast, direct answer. Minimal reasoning.
• Effort 4-6: Balanced. Moderate depth for standard problems.
• Effort 7-10: In-depth. Comprehensive analysis, many steps.
⚡
Fast Response
At low effort values (1-3), Claude responds immediately, perfect for real-time applications.
🧠
Deep Reasoning
At high effort values (8-10), Claude performs detailed analyses for complex problems.
💰
Cost Optimization
With effort control, you automatically optimize between quality and API costs.
🎯
User Control
Unlike hidden reasoning: You decide the reasoning intensity, not the model.