What Is a Token in AI?
Send GPT or Claude what looks like a short prompt, and sometimes the bill that comes back doesn't match. There's a specific reason for that: tokens. An AI model doesn't read your text the way you do — it reads tokens, small chunks somewhere between a letter and a full word, and that's what you're actually paying for. Once you see how the counting works, the pricing stops feeling random.
How Tokens Work
Before an AI model does anything resembling "thinking," it runs your text through a tokenizer, which chops it into pieces. Each piece is a token. For English, that's roughly four characters per chunk — about three-quarters of a word on average. The real number, though, depends entirely on what you typed.
"Hello world" is a clean example: two tokens, one per word. Nothing surprising there. Things get messier with longer or rarer words. "Tokenization" almost never survives as a single piece — most tokenizers split it into "token" and "ization." Typos, made-up words, names the model rarely saw during training — they all get fragmented the same way, simply because nothing was reserved for them in advance. Even numbers are inconsistent: "2026" might stay whole, or it might split into "20" and "26" depending on the model. Small thing, but it's exactly the kind of detail that makes counting by hand a losing game.
Punctuation counts too, just cheaply. A period or comma is usually its own unit, while a leading space tends to get folded into whatever word follows rather than standing on its own. Underneath all of it sits a compression algorithm — most major models run some flavor of byte-pair encoding — built to represent common English efficiently while still being flexible enough to handle whatever gets thrown at it.
Why Tokens Matter for Cost
Here's where it actually hits your wallet: OpenAI, Anthropic, Google — they all bill per token, not per word, not per character. Two prompts that look about the same length on screen can end up costing noticeably different amounts, purely based on what's in them. Plain, everyday words compress well. Jargon, code, or anything unusual doesn't.
I'd also flag this, because it surprises people: a prompt written in Hindi, Arabic, or Japanese will usually eat more tokens than the same sentence in English. Most tokenizers were trained on English-heavy data, so other scripts get cut into smaller, less efficient pieces. Code has the same problem — a few lines of Python can cost more than a full paragraph of prose, since brackets and indentation don't map neatly onto the shortcuts built for ordinary words.
None of this is something you're meant to calculate by hand. People who try usually land somewhere close with "characters divided by four," but close isn't the actual number — it ignores which words you actually used. That gap is the whole reason a token counting tool exists instead of everyone eyeballing it.
Input vs Output Tokens
Tokens aren't priced evenly. Providers split the bill into input (what you send) and output (what the model writes back), and output almost always costs more — sometimes several times more.
Why? Reading your prompt is a light, single pass for the model. Writing a reply is a different kind of work — predicting one token at a time, each one shaped by everything generated before it. That's heavier, computationally, and the price reflects it.
Claude Sonnet 5 is a good example: about $2 per million input tokens, $10 per million output — five times the rate. On one short exchange, that gap is a fraction of a cent. Run thousands of requests a day, or ask for long, detailed replies instead of short ones, and it adds up fast. A short question with a long answer can end up pricier than a long question with a short one, which catches a lot of people off guard the first time they actually look at their usage.
See the exact token count and cost for your own prompt, live, as you type.
Try the calculator →Frequently Asked Questions
How many words is one AI token?
Slightly less than one. English averages around four characters per token, so a 1,000-word page usually lands around 1,200–1,300 tokens — not an even 1:1.
Do GPT and Claude count differently?
Yes, more than most people expect. Each provider runs its own tokenizer, so identical text can produce different counts across GPT, Claude, and Gemini. There's no single "correct" count — it depends entirely on which model is reading it.
Do spaces and punctuation count?
Mostly, yes, but cheaply. A comma or period is usually its own token; a leading space often just attaches to the next word instead of standing alone.
Does a longer prompt always cost more?
Not always. Character count is a rough guide at best. A short prompt loaded with rare words or code can out-cost a longer one written in plain English.
Check Your Real Number
Estimating gets you in the ballpark, but when actual money's about to leave your account, "in the ballpark" isn't quite enough. Our free AI token calculator shows the exact count as you type, right in your browser — nothing gets sent anywhere.