What Are AI Tokens? A Plain-Language Guide

Every AI LLM — ChatGPT, Claude, Gemini — reads your prompt in small pieces called tokens. A token is roughly a short word fragment: “cat” is one token, “strawberry” is two or three, and a long technical word can shatter into five or more. The model never sees your words; it sees these numbered fragments.

Why tokens exist

Language models process text by mapping each fragment to an ID number in a fixed vocabulary — often 50,000 to 100,000 entries. Training and prediction happen on those IDs. That’s fast, but it means your prompt’s cost and its processing time both scale with token count, not with your idea of its “length.”

The 4-character rule of thumb

For English, a token averages three to four characters. A 1,000-word prompt is roughly 1,300 to 1,600 tokens. Punctuation, numbers, and unusual spellings all shift the count, and every model family uses its own tokenizer, so the same text tokenizes differently across vendors.

That’s why any token counter — including the one in our compressor — is an estimate, not an invoice. It’s honest enough to make decisions with, and that’s what it’s for.

Where tokens touch your wallet

• API billing prices per million tokens, in and out.
• Context windows are token budgets: a 128k-token model can’t hold a 140k-token conversation, no matter how good it is.
• Rate limits on LLM subscriptions count tokens per minute or per day.

The takeaway

If you use AI LLMs seriously, tokens are your unit of spend — like kilowatt-hours on an electricity bill. And just like a power bill, the first saving isn’t a better rate: it’s using less. The cheapest token is the one you never send.

Strip the filler and pleasantries from your prompts with supatoken.lol — it’s free, runs in your browser, and nothing you paste is stored.