What is a Token?

Why tokens exist

Computers don’t read letters the way you do. Before an AI model can process your message, it needs to convert your text into numbers. Tokenization is that conversion: breaking text into pieces the model can work with.

The word “hello” might be one token. The word “tokenization” might be three tokens: “token”, “ization”. A space before a word often gets included in the token. Punctuation is usually its own token.

The exact split depends on the tokenizer, and different AI models use different ones. But the rough math holds for English: 1 token ≈ 4 characters ≈ ¾ of a word. Other languages, code, and newer tokenizers can differ noticeably. Anthropic, for example, says the tokenizer introduced with Claude Opus 4.7 produces roughly 30% more tokens for the same text than its previous one, so the same prompt can cost a different number of tokens on different models.

Why this matters for your wallet

AI companies charge per token. Here are some list prices at the time of writing (September 2026, standard API rates; prices change often, see the LLM Landscape 2026 for the current list):

Gemini 3.8 Flash’s price is introductory until 31 December 2026 and doubles to $1.50 / $7.50 from 1 January 2027. OpenAI charges GPT-6 prompts above 272K tokens at 2× the input and 1.5× the output rate. The spread between the cheapest and the most expensive row is roughly 40× on both input and output, which is why picking the right model matters more than almost anything else for cost.

…this is what they mean. You pay for:

• Input tokens: Everything you send to the model (your question, any documents you paste, the system prompt)
• Output tokens: Everything the model generates back (usually more expensive because generation is harder than reading). On reasoning models, the hidden “thinking” the model does before answering is billed as output tokens too, so a short visible answer can still cost thousands of output tokens.

A typical short conversation might use 500 input tokens and 300 output tokens. That’s fractions of a cent. But if you’re building an app that sends a 10-page document to Claude every time a user asks a question, and you have 1,000 users per day, the math changes fast.

Quick token math

• 1 page of text ≈ 500 tokens
• A typical email ≈ 200-400 tokens
• “What’s the weather?” ≈ 5 tokens
• A detailed system prompt ≈ 500-2,000 tokens
• An entire novel ≈ 100,000+ tokens

You can check exact token counts using OpenAI’s tokenizer tool or Anthropic’s token counter in the API response.

The context window limit

Every model has a maximum number of tokens it can handle at once, called the context window:

The window is shared between input and output, and most models also cap the output separately. If you’re at 195,000 tokens on a 200,000-token model and ask for a 10,000 token response, you’ll hit the wall. A bigger window is not free either: every token you send is billed, and some providers charge more per token for very long prompts.

When you exceed the context window:

• The API returns an error
• Or the model “forgets” the oldest messages to make room
• Or your request just fails silently

This is why apps that work with documents need strategies like chunking (splitting documents into pieces) or RAG (only retrieving relevant sections).

Why your vibe coding costs add up

Common token traps for vibecoders:

Huge system prompts: If you paste your entire codebase into the system prompt “for context,” you’re paying for that on every single message.

Asking for long responses: “Explain everything in detail” costs more than “Give me a one-sentence summary.”

Using expensive models for simple tasks: A flagship model like Claude Opus 5.5 or GPT-6 Sol is overkill for “Is this email spam? Yes or no.” A smaller, cheaper model such as GPT-6 Luna, Gemini 3.8 Flash or Claude Haiku 4.5 works fine, at a fraction of the price.

Not caching: If 10,000 users ask the same question, you pay for the AI to answer it 10,000 times. Caching identical responses saves money. Providers also offer prompt caching: if every request starts with the same long system prompt or document, the repeated part is billed at a steep discount (for example $0.20 instead of $2 per 1M input tokens on GPT-6 Sol, and $0.20 instead of $4 on Claude Opus 5.5).

What to do about it

1. Know your costs: Check your AI provider’s dashboard. See what’s actually costing money.

2. Right-size your model: Use cheap models for simple tasks, expensive models for hard ones.

3. Be specific: “Summarize in 2 sentences” costs less than “Tell me everything about this.”

4. Cache when possible: If the same input always produces the same output, cache it.

5. Use retrieval: Don’t send entire documents. Retrieve only the relevant chunks.

Further reading

• What is an LLM? : The model that consumes these tokens
• LLM Landscape 2026 : Current models, prices and context windows
• What is RAG? : How to work with documents without blowing your token budget
• Context engineering : The discipline of managing what goes into the context window
• Tokenization : The technical details of how text becomes tokens
• What is JSON? : The data format you’ll see in API responses about token usage
• What is an API key? : Required to access AI APIs that charge per token
• What is rate limiting? : Token-based rate limits (TPM) alongside request limits
• Tokenmaxxing : The organizational pattern of maximizing token usage as a productivity signal