Tokens Explained: How AI Actually Reads Your Words

When you type a sentence into ChatGPT, Claude, or any other AI chatbot, it doesn't see your words the way you do. It doesn't read letter by letter either. It reads something in between, called a token. Tokens are the quiet unit behind almost every surprising thing AI does: how much you can paste in at once, how much it costs the company running the model, and why a tool that can write a decent essay sometimes can't tell you how many r's are in "strawberry."

Once you understand tokens, a lot of AI's weird behavior stops being mysterious.

The analogy: puzzle pieces, not letters

Imagine the AI has a box of about 100,000 puzzle pieces. Each piece has a chunk of text printed on it. Some pieces are whole common words like "the" or "because." Some are fragments like "ing" or "tion." Some are a single character. Rare words get broken into several pieces. Common words get their own piece.

When you send a message, the model's first job is to find the pieces that match your text and lay them out in order. It doesn't see "unbelievable." It sees something like three pieces: "un" + "believ" + "able." When it writes back, it's picking the next puzzle piece, over and over, based on everything it has seen so far.

That's a token. A chunk of text, usually smaller than a word, that the model treats as one unit.

A rough rule of thumb

For English, one token is about four characters, or roughly three quarters of a word. So:

• 100 tokens is about 75 words, or a short paragraph.
• 1,000 tokens is about 750 words, or a page.
• 100,000 tokens is a short novel.

Other languages, code, and unusual words use more tokens per word. A document in Japanese or a block of JSON will "cost" more tokens than the same amount of plain English.

Why tokens show up on your bill and in length limits

Every AI provider charges by the token, both for what you send in and what the model writes back. When you see a plan that says "1 million tokens included," that's a budget of puzzle pieces, roughly 750,000 words of combined input and output.

Every model also has a context window, which is the maximum number of tokens it can hold in its head at once. If a model's window is 200,000 tokens, you can paste in about 150,000 words of material and still leave room for its reply. Go over, and the oldest material falls off the back, or the system refuses the request.

This is why pasting an entire 400-page PDF sometimes works and sometimes doesn't, and why asking a chatbot to "remember" something from two hours ago often fails. There is a hard ceiling on how many puzzle pieces fit on the table.

Why AI is strangely bad at spelling and counting letters

Here is the fun part. If the model sees "strawberry" as two tokens, something like "straw" + "berry," then the individual letters are not really visible to it. It knows the word, knows what it means, knows how to use it in a sentence. But asking "how many r's are in strawberry" is like asking someone to count the sesame seeds on a hamburger bun they've only ever seen wrapped.

The model has to reason its way to the answer from the outside, and it often slips. This is also why AI can confidently misspell a name, reverse letters in a short word, or fail at a simple anagram. Not because it's dumb, but because it was never looking at letters in the first place.

Newer models are better at this, partly because they've been trained to "spell out" the word internally before answering. But the underlying quirk is real: the native unit of thought is the token, not the character.

What this means for how you use AI

A few practical takeaways:

• Big pastes have a ceiling. If you're dumping a long transcript, email thread, or document in, check whether the tool tells you how close you are to the window. If it starts "forgetting" the top, that's why.
• Shorter prompts aren't always better, but padded ones aren't free. Every "please" and "thank you" is tokens. It won't break your budget, but on heavy automated use it adds up.
• Don't trust it for letter-level tasks. Counting letters, finding every word that starts with "s," or exact-character edits are the kinds of jobs where AI quietly misses. Verify by hand, or ask it to write code that does the counting.
• Non-English text costs more. If you're translating or working in another language, expect to hit limits faster.

Tokens are one of those invisible plumbing details that explain more about AI than any marketing page will. Once you can see them, the quirks start making sense, and you stop fighting the tool at the places where it was never going to win.

If posts like this are useful, you can get the next one by subscribing on the NeuroByte blog.