glossary
What Are AI Tokens? How AI Billing Works
A token is a chunk of text — usually 3–4 characters — that AI models read and generate one piece at a time. Roughly 1,000 tokens equal 750 words. Tokens are the unit API pricing and context windows are measured in, so they quietly shape every AI bill.
4 min read · Updated Apr 17, 2026
A token is a piece of a word, not a whole word
Language models don’t read whole sentences at once — they read in chunks called tokens. A token is usually 3 to 4 characters, so a common word like “cat” is one token, “running” might be two, and a rare word might be split into several.
The quick rule: 1,000 tokens ≈ 750 words of English. Multiply by 1.3 and you have a rough word-to-token conversion.
Why tokens exist
AI models work with a fixed vocabulary of tokens (typically 100K to 200K unique pieces). Breaking text into tokens lets the model handle any word — even ones it’s never seen — by combining pieces it already knows.
This is also why tokenization varies across providers: each company trains its own tokenizer with different choices about how to split words.
Where you’ll bump into token counts
- Context window. A 128K-token window fits about 96,000 words of conversation, system prompt, files, and response combined.
- Message limits. Plans like ChatGPT Plus don’t publicize token caps, but very long chats count more tokens per turn and push you toward rate limits faster.
- API pricing. If you use the API (not consumer chat), everything is priced per million tokens — input cheaper, output more expensive.
- File uploads. A PDF isn’t free; it’s tokenized before the model reads it, and those tokens come out of your context budget.
Why this matters for picking a subscription
On the surface, consumer subscriptions look flat-fee: pay $20/month, chat all you want. Under the hood, every plan has a token ceiling — the context window, the message cap, the file upload size. Understanding tokens helps you:
- Choose a plan with enough window for your real work (see Context Window).
- Stop wondering why ChatGPT Plus sometimes truncates a long document.
- Decide when a 1M-token Claude plan or 2M Gemini plan is actually worth the upgrade.
Tokens in different languages
English is roughly the most efficient. Chinese, Japanese, Korean, and languages with long compound words use more tokens per word. If you work in another language, your effective context window is smaller than the headline number.
Going deeper
Frequently asked questions
How many words are in 1,000 tokens?
About 750 English words. The ratio varies with language and formatting — code, numbers, and non-English text can use more tokens per word.
Do consumer subscriptions charge per token?
No. Subscriptions like ChatGPT Plus and Claude Pro charge a flat monthly fee with message limits. Token pricing applies to API usage, not the consumer chatbots.
Why do tokens matter if I'm on a flat-fee plan?
Tokens still determine your context window size and message limits. If you paste a 50-page document, you're spending tokens — and hitting the window limit faster — even if no invoice changes.
How does a model turn a word into a token?
A tokenizer splits text using byte-pair encoding or a similar scheme. 'ChatGPT' might be one token; 'unbelievable' might be two; rare words or emoji can span multiple tokens.
Is one token the same across ChatGPT, Claude, and Gemini?
No. Each provider uses its own tokenizer, so the same sentence might be 18 tokens in one and 22 in another. The 1,000 ≈ 750 words rule is a rough average.
What's an 'output token' vs an 'input token'?
Input tokens are what you send the model (your prompt, chat history, files). Output tokens are what the model writes back. On API pricing, output tokens usually cost more than input tokens.