← Back to blog

How to Estimate Token Usage for Your AI Application

Tokens are the currency of LLM APIs — every request costs you input and output tokens, and every token costs money. But most developers have no idea how many tokens their app actually uses until the bill arrives. This guide gives you a practical framework for estimating token usage before you ship.

What Is a Token?

A token is roughly 4 characters of English text, or about 3/4 of a word. But it's not exact — tokenization depends on the tokenizer used by each model. Here are some practical benchmarks:

Content	Approximate Tokens	Rule of Thumb
1 word	~1.3 tokens	Words × 1.3
1 sentence (15 words)	~20 tokens	Sentences × 20
1 paragraph (100 words)	~130 tokens	Paragraphs × 130
1 page (500 words)	~650 tokens	Pages × 650
1,000 words	~1,300 tokens	Words × 1.3
1 code file (100 lines)	~500-800 tokens	Lines × 5-8
1 JSON object (50 keys)	~300-400 tokens	Keys × 6-8

Key insight: Tokens aren't words. A 1,000-word document is roughly 1,300 tokens. Code is denser than prose — each line of code is typically 5-8 tokens depending on complexity.

Token Estimation Rules by Use Case

Chatbot / Conversational AI

Rule of thumb: Budget 1,500 tokens per chatbot request (1,000 input + 500 output). If you maintain conversation history, add 200 tokens per previous turn.

Code Generation

Rule of thumb: Budget 1,500 tokens per code generation request. Code completion requests (autocomplete) are lighter — around 500-800 tokens total.

Document Analysis / Summarization

Rule of thumb: Document analysis is input-heavy. Budget 3,000 tokens per request (2,500 input + 500 output). For longer documents, use the formula: pages × 650 + 800.

RAG (Retrieval-Augmented Generation)

Rule of thumb: Budget 1,800 tokens per RAG request (1,400 input + 400 output). The key variable is chunk size — larger chunks mean more input tokens but potentially better answers.

Classification / Extraction

Rule of thumb: Classification is the cheapest use case. Budget 500 tokens per request (450 input + 50 output). Use cheaper models like GPT-4o mini or Gemini Flash for classification.

Step-by-Step Estimation Framework

1. Map your API calls: List every place your app calls an LLM API. Include the use case, model, and trigger frequency.
2. Estimate tokens per call: Use the rules of thumb above. When in doubt, overestimate by 20%.
3. Estimate daily volume: How many requests per day? Multiply by tokens per call for daily token usage.
4. Separate input and output: Most providers charge differently for input vs output tokens. Split your estimate.
5. Calculate monthly totals: Daily tokens × 30 = monthly tokens. Convert to millions (divide by 1,000,000) for pricing.
6. Apply pricing: Multiply input tokens (in millions) by input price, output tokens by output price. Add them up.

Worked Example: AI Code Review Bot

With GPT-4o mini, the same bot costs just $1.47/month. Model selection is the single biggest factor in your API bill.

Common Estimation Mistakes

• Forgetting system prompts: System prompts are sent with every request. A 500-token system prompt × 10,000 requests = 5M tokens/month you didn't plan for.
• Ignoring conversation history: Chatbots send the full conversation with each turn. A 10-turn conversation grows from 1K to 10K+ input tokens.
• Underestimating output tokens: Set max_tokens limits. Without them, models can generate 4K+ tokens per response.
• Not accounting for retries: Failed requests still cost tokens. Budget 5-10% extra for retries and errors.
• Using word count instead of token count: Remember: tokens ≠ words. Always multiply word count by ~1.3.

Token Optimization Tips

1. Trim conversation history: Keep only the last 5-10 turns instead of the full conversation.
2. Set max_tokens: Cap output tokens to what you actually need. A 200-token response doesn't need a 4,096 limit.
3. Use shorter system prompts: Every token in the system prompt is multiplied by every request. Keep it concise.
4. Batch similar requests: Combine multiple classification tasks into a single prompt when possible.
5. Choose the right model: Use GPT-4o mini or Gemini Flash for simple tasks. Save premium models for complex reasoning.
6. Cache repeated prompts: If you send the same prompt frequently, use provider caching (OpenAI, Anthropic both offer this).