Words are human units
Words are useful for reading and writing, but they are not the unit most AI APIs use for billing.
Tokens vs words
Tokens vs Words
Words are easy for humans to count, but AI models usually process text as tokens. Paste your text to compare characters, words, estimated tokens, and potential cost.
Calculator
Estimate your AI prompt cost
Paste a prompt, choose an example pricing profile, and estimate cost per prompt run, per day, and per month.
Quick examples
Input tokens are what you send to the AI model. Output tokens are what the model returns. API providers often price them separately.
Advanced settings
Prices are manual for now. Example: if your provider charges $2 input and $10 output per 1M tokens, enter 2 and 10.
Energy usage is a rough estimate. Actual energy depends on model, hardware, provider, datacenter efficiency, workload, and region.
Words are not billing units
A 500-word article, a 500-word JSON sample, and a 500-word code block can produce different token counts and different costs.
Tokens are model units
A token can be a word, part of a word, punctuation, whitespace, code, or formatting.
Counts vary by text
Language, punctuation, code, markdown, JSON, and message structure can all change token estimates.
Cost usually follows tokens
AI cost is usually based on input and output tokens, not word count. PromptMeter estimates all three side by side.
Content type guide
Simple prose is often easier to estimate. Technical text, code, JSON, and markdown can use more tokens because symbols and structure count too.
Language guide
English, Spanish, German, Chinese, and Japanese can tokenize differently. Treat every language-specific estimate as approximate.
Content type orientation
| Content type | Token behavior | Notes |
|---|---|---|
| Simple prose | Usually close to the general estimate | Varies by language and punctuation |
| Technical text | Often slightly denser | Acronyms and symbols can change counts |
| Code | Often denser | Brackets, operators and indentation matter |
| JSON | Often denser | Keys, quotes and repeated structure add tokens |
| Markdown | Variable | Lists, headings and formatting affect estimates |
These are orientation tables, not official tokenizer measurements.
Language and script orientation
| Language/script | Why it can vary |
|---|---|
| English | Often close to common token estimates |
| Spanish/French/Italian/Portuguese | Accents, longer words and punctuation can shift estimates |
| German/Dutch/Polish/Russian | Compound words and morphology can change token counts |
| Chinese/Japanese/Korean | Character-based scripts behave differently from word-based estimates |
| Code/structured text | Structure can matter more than natural language |
These are orientation tables, not official tokenizer measurements.
Why word count can mislead
| Text type | Why word count can mislead | Better estimate |
|---|---|---|
| Plain prose | Words may map roughly to common token estimates, but punctuation and language still matter | Estimate tokens directly |
| Code | Operators, brackets, indentation, and short identifiers count even when word count is low | Use a token estimate with code-heavy assumptions |
| JSON | Keys, quotes, braces, commas, and repeated structure add tokens | Estimate input and output tokens separately |
| Markdown | Headings, lists, links, and tables add formatting tokens | Compare characters, words, and token estimates |
| Long answers | Billing depends on generated tokens, not the words you originally sent | Use an output-token cost estimate |
Cost depends on input and output tokens, not word count alone.
FAQ
Tokens vs words FAQ
Is one word the same as one token?
No. Some words are one token, some split into multiple tokens, and punctuation or formatting can count too.
Why do tokens matter for AI cost?
Providers often price API usage by tokens. More input or output tokens usually means higher cost.
Can token counts vary by model?
Yes. Different models and tokenizers can count the same text differently, so these estimates stay approximate.
Does JSON count differently from prose?
Often yes. Keys, punctuation, braces, brackets, indentation, and repeated fields can increase token density.