CHAPTER 8.5b · TOKEN ECONOMY & COSTS

Token Economy Calculator

Calculate the real cost per request based on input/output tokens and model pricing

The Token Economy Calculator shows what a single API call costs – and how it adds up at 1,000 or 1 million requests. Essential for budget planning and model selection.

📖 Learning Context ▼
🎯

Learning Objectives

  • Calculate cost per request
  • Distinguish input vs. output token costs
  • Compare ROI for different models
🗺️

Context: Where are we?

Step 3/5 Chapter 8: Tools & Glossary

Calculators (3/5) enable informed decisions when choosing models and providers.

💡

Why It Matters

A chat request can cost $0.001 or $0.10 – a 100x difference! Understanding token economics means choosing the right model for the use case.

📌

Key Takeaways

  • $/1M Tokens: Standard metric for comparisons
  • GPT-4 vs. Haiku: ~50x price difference
  • Break-even: When does fine-tuning pay off?
Input Cost
$0.0075
Output Cost
$0.0060
Total Cost
$0.0135
Per 1M Tokens
$24.29
Model Input ($/1M) Output ($/1M) Type Use Case
GPT-4 $15 $30 Expensive Complex Reasoning, Code
Claude 3.5 $3 $15 Medium Balanced, General
Llama 3.1 405B $2.70 $8.10 Affordable Open-Source, Commodity
DeepSeek $0.14 $0.28 Ultra-cheap Bulk Processing, Fine-tuning
💰
Output is More Expensive Than Input
GPT-4: Input $15, Output $30 (2x). DeepSeek: 2x as well. Longer outputs = proportionally higher costs. Significant at 1000 output tokens.
📊
Token Counting: ~4 chars = 1 Token
1000 words ≈ 1500 tokens. "Hello" = 1 token, but "ChatGPT" = 1-2. Models count differently (BPE variants). Always test!
🌍
DeepSeek Price Revolution
DeepSeek: $0.14/$0.28 (100x cheaper than GPT-4). Makes bulk processing & fine-tuning economical. But quality trade-offs on complex tasks.
✅
Claude 3.5 Best Value
$3/$15 (5x cheaper than GPT-4, better quality than Llama). For production: Claude or Llama, for cost: DeepSeek, for quality: GPT-4/o3.
🔧
Self-Hosted = High Upfront
Llama on A100: $1-5 per million tokens (infra costs). Worth it from ~100M tokens/month. Below that: API is better.
📈
Prices Falling Continuously
GPT-4 prices halved in 2024. DeepSeek undercut everyone in 2025. Trend: Commodity prices for models <100B parameters.