What Is a Token? How AI Coding Tools Count and Bill Tokens (2026 Guide)

The Short Answer

A token is the unit an AI model uses to read and write text. It's not quite a word and not quite a character — it's a chunk of text, usually a few characters long. Every AI coding tool bills you by the token, so understanding what a token is gives you direct insight into why your AI bill looks the way it does.

A useful rule of thumb: 1 token ≈ 4 characters of English text, or roughly ¾ of a word. So 1,000 tokens is about 750 words. Code tends to tokenize a little differently — punctuation, indentation, and symbols each cost tokens — but the 4-characters-per-token estimate is close enough for budgeting.

Why Models Use Tokens Instead of Words

Models don't read characters or whole words — they read tokens, which are produced by a process called tokenization. A tokenizer breaks text into common sub-word pieces. Frequent words like "the" become a single token; rarer or compound words get split into several. The word "tokenization" might become "token" + "ization," for example.

This matters for code because source code is full of things that don't tokenize neatly. A variable named getUserAccountBalance may split into multiple tokens. Whitespace, braces, and semicolons each consume tokens. That's why a 100-line file can be more tokens than you'd guess from its word count alone — code is denser in tokens than prose.

How Tokens Become Your Bill

AI providers price in dollars per million tokens, and they charge separately for input tokens (what you send — your prompt, your code, the conversation history) and output tokens (what the model generates back). Output almost always costs more than input.

Here's how a few 2026 models price it, per million tokens (input / output):

• Claude Opus 4.8: $5 / $25
• GPT-5.5: $5 / $30
• Claude Sonnet 4.6: $3 / $15
• Gemini 3 Flash: $0.50 / $3
• DeepSeek V4 Pro: $0.435 / $0.87

Notice the spread: the most expensive output rate (GPT-5.5 at $30/M) is nearly 35× the cheapest here (DeepSeek V4 Pro at $0.87/M). The model you pick moves your bill far more than any other single factor.

A Worked Example

Say you ask an AI coding tool to refactor a file. You send 6,000 tokens of context (the file plus instructions) and it returns 2,000 tokens of revised code. That's 6K input + 2K output.

With Claude Opus 4.8: 6,000 × $5/M = $0.03, plus 2,000 × $25/M = $0.05, for $0.08 total. With DeepSeek V4 Pro: 6,000 × $0.435/M + 2,000 × $0.87/M ≈ $0.004. Same task, a 20× cost difference, driven entirely by token pricing.

The catch is that this counts one clean pass. Real coding sessions involve back-and-forth, and each turn re-sends the growing conversation as input. That's why long sessions get expensive — the input token count climbs with every exchange.

What This Means for Your Budget

Once you think in tokens, the levers become obvious. Send less context — only the files that matter, not your whole repo. Use prompt caching so repeated context isn't billed at full input price every turn. Match the model to the task — a cheap model for boilerplate, a frontier model only where it earns its rate.

Understanding tokens turns your AI bill from a mystery into arithmetic. To put real numbers on your own project, plug your expected usage into our AI cost calculator and compare models side by side.