Skip to content
AYOCORP Logo AYOCORP Logo AYOCORP
Blog Docs

Tokens

An illustration of planets and stars featuring the word "astro"

Tokens are how AI agents “read” text. Understanding tokens helps you write better instructions and predict costs.

What

Token = AI’s Unit of Text

AI doesn’t read words like humans. It breaks text into “tokens” - chunks that can be whole words, parts of words, or punctuation.

Examples:

  • “Hello” = 1 token
  • “Hello world” = 2 tokens
  • “Hello, world!” = 4 tokens (Hello, ,, world, !)
  • “AI agent” = 2 tokens
  • “understanding” = 2 tokens (under, standing)

Why it matters: AI providers charge by tokens, not words. Longer instructions = higher costs.

How It Works

What happens when you use AI:

  1. Your text becomes numbers

    • You write: “Why is the moon light”
    • AI sees: [101774, 623, 28157, 6038]
    • Each number represents a word or word-part

  2. AI thinks in numbers

    • Does calculations on those numbers
    • Figures out the best response

  3. Numbers become text again

    • AI’s number output converts back to words
    • You see readable text

Simple example:

  • You send 2,300 characters of text
  • AI converts it to 484 tokens (numbers)
  • You get billed for 484 tokens, not 2,300 characters

Why you should care: Tokens are what you actually pay for. A long email might use fewer tokens than you think, which means lower costs than expected.

Business

Real-world cost examples using current AI model pricing:

Email Processing Task

Using Claude Sonnet 4.5 (High-quality responses)

  • Task: Categorize and respond to 100 emails daily
  • Tokens per email: ~500 input + ~200 output = 700 tokens total
  • Monthly volume: 100 emails × 30 days = 3,000 emails
  • Monthly cost:
    • Input: 1.5M tokens × $3.00 = $4.50
    • Output: 600K tokens × $15.00 = $9.00

Using GPT-5 (Balanced cost and quality)

  • Same task: 100 emails daily
  • Monthly cost:
    • Input: 1.5M tokens × $1.25 = $1.88
    • Output: 600K tokens × $10.00 = $6.00

Compare to: Administrative assistant at $3,000-4,000/month

Document Processing Task

Using Claude Sonnet 4.5

  • Task: Summarize 50 business reports monthly
  • Tokens per report: ~2,000 input + ~300 output
  • Monthly cost:
    • Input: 100K tokens × $3.00 = $0.30
    • Output: 15K tokens × $15.00 = $0.23

Using GPT-5

  • Same task: 50 reports monthly
  • Monthly cost:
    • Input: 100K tokens × $1.25 = $0.13
    • Output: 15K tokens × $10.00 = $0.15

Compare to: Analyst spending 10-15 hours on this work

Data Extraction Task

Using Claude Sonnet 4.5

  • Task: Extract data from 50 invoices daily
  • Tokens per invoice: ~600 input + ~100 output
  • Monthly cost:
    • Input: 900K tokens × $3.00 = $2.70
    • Output: 150K tokens × $15.00 = $2.25

Using GPT-5

  • Same task: 50 invoices daily
  • Monthly cost:
    • Input: 900K tokens × $1.25 = $1.13
    • Output: 150K tokens × $10.00 = $1.50

Compare to: Data entry clerk at $2,500-3,500/month

The bottom line: AI automation dramatically reduces operational costs for routine tasks, freeing up human resources for higher-value strategic work.

Pricing source: models.dev (November 2024). Costs are per 1 million tokens.

Counting

Rough token estimates:

  • 1 word = 1.3 tokens on average
  • 1 page of text = ~750-1,000 tokens
  • 1 email = ~200-800 tokens
  • 1 business document = ~1,000-5,000 tokens

Quick calculation:

Your text length ÷ 0.75 = approximate tokens

Example: 300 words ÷ 0.75 = ~400 tokens

Optimization

Save tokens (and money) with these tips:

Shorter Instructions

❌ Verbose:

Please read through this customer email carefully and analyze
the content to determine what type of inquiry this represents,
considering whether it might be a sales question, a support
request, a billing issue, or some other category, and then
provide your assessment.

✅ Concise:

Categorize this email: sales, support, billing, or other.

Remove Fluff

❌ Unnecessary words:

I would like you to please extract the following information
if possible from this document: customer name, order amount,
and delivery date. Thank you.

✅ Direct:

Extract: customer name, order amount, delivery date.

Use Templates

❌ Repeating context:

You are a customer service agent for ABC Company. Read this
email and respond professionally...
[repeated for every email]

✅ Reference template:

Use customer service template. Respond to this email.

Monitoring

Track your token usage:

  1. Daily tracking - Most platforms show token consumption
  2. Cost alerts - Set spending limits ($10-50/month typical)
  3. Usage patterns - Which tasks use most tokens?
  4. Optimization opportunities - Where can you shorten instructions?

Typical business costs:

  • Small business: $5-25/month
  • Growing company: $25-100/month
  • Enterprise: $100-500/month

Limits

AI agents have token limits per conversation:

Standard Models

Limit: ~4,000-8,000 tokens Handles: Most business documents Best for: Emails, forms, short reports

Extended Models

Limit: ~32,000-128,000 tokens Handles: Long documents, multiple files Best for: Contracts, research reports, manuals

What happens at the limit:

  • Agent stops processing
  • Only sees first part of your document
  • May give incomplete results

Solution: Break large documents into smaller chunks.

Practical

Real scenarios and token impact:

Daily Email Processing

Instruction: 50 tokens
Average email: 300 tokens
Response: 100 tokens
Total per email: 450 tokens
100 emails daily = 45,000 tokens = ~$0.09/day

Invoice Data Extraction

Instruction: 30 tokens
Invoice text: 400 tokens
Data output: 50 tokens
Total per invoice: 480 tokens
50 invoices daily = 24,000 tokens = ~$0.05/day

Weekly Report Generation

Source data: 2,000 tokens
Instruction: 40 tokens
Generated report: 500 tokens
Total per report: 2,540 tokens
4 reports monthly = 10,160 tokens = ~$0.02/month

Management

Control token costs:

Set Budgets:

  • Start with $20/month limit
  • Monitor usage first 2 weeks
  • Adjust based on actual needs

Optimize Instructions:

  • Remove unnecessary words
  • Use abbreviations where clear
  • Reference previous context instead of repeating

Batch Processing:

  • Process multiple items together
  • Share context across similar tasks
  • Use templates for repeated instructions

Common

“This seems expensive…” Compare to human time. 1,000 tokens costs ~$0.002. Human time for same task: $5-20.

“How do I count tokens before sending?” Most platforms show estimates. For planning, use: word count ÷ 0.75.

“What if I hit the limit?” Break tasks into smaller pieces. Process documents in sections.

“Do I pay for bad responses?” Yes, you pay for the attempt. That’s why good instructions save money.

Start

Right now:

  1. Check your current usage on your AI platform
  2. Calculate cost per task for your main use case
  3. Set a monthly budget ($20-50 for most businesses)
  4. Optimize one instruction to use fewer tokens

Next week:

  • Review actual costs vs estimates
  • Identify your highest-token tasks
  • Look for optimization opportunities

Next

Explore related concepts:

  • Prompting - Write efficient instructions that save tokens
  • Models - Understand how models generate responses
  • Data - Organize data to reduce token usage

Understanding tokens helps you optimize costs.