ChatGPT Token Counter

Language models do not read word by word. They break text into tokens, which are chunks of characters produced by an algorithm called byte pair encoding. A token is roughly three quarters of an English word on average, but that average hides a lot of variation.

Most token counter pages tell you they are accurate and leave it there. Here is the actual position.

Real tokenisation needs the model's full vocabulary file, which is several megabytes. Loading that on every page view would make this page slow for everyone, so like every browser-based counter this one estimates. It does the same first step OpenAI's tokenizer does, splitting text on word, digit, punctuation and whitespace boundaries, then estimates how each piece would be broken down further.

Where it is reliable

Ordinary English prose. Common words are single tokens in the real tokenizer and single tokens here, so a normal paragraph usually lands within a token or two of the true count.

Where it drifts

Long technical words, compound words, unusual names and dense code. The real tokenizer splits those into several pieces, and any estimator has to guess where. Expect a difference of a token or two on that kind of text, occasionally more.

Claude and Gemini use different tokenizers again, and neither is published in a form a browser can run. Those counts apply a rough ratio to the GPT estimate, so treat them as a ballpark rather than a number to budget against.

1. Paste your text into the box above. The count updates as you type.
2. Pick the model you are actually using, since context limits and prices differ enormously.
3. Set the output tokens you expect the reply to be, because on most models output costs several times more than input.
4. Check the context bar. If it is climbing past 80 percent you are close to the point where the model starts forgetting.

Use the System + User tab for API work

A system prompt is sent again with every single request. A 400 token system prompt across ten thousand calls is four million input tokens you may not have counted. Splitting them out shows you which half of your prompt is the recurring cost.

Use Compare Models before you commit

The same prompt can differ by a factor of fifty in price between the cheapest and most expensive model. The compare tab sorts every model cheapest first and flags any whose context window your prompt would not fit into.

API pricing is quoted per million tokens, and input and output are charged at different rates. The arithmetic is simple:

Two things catch people out. First, output is usually three to five times the price of input, so a chatty model that answers at length costs far more than the prompt suggests. Second, in a conversation the entire history is resent as input on every turn, so a long chat costs more per message as it goes on, even if your messages stay short.

The context window is the total number of tokens a model can hold at once: your system prompt, the whole conversation so far, and the reply it is about to write, all together.

When you exceed it, nothing dramatic happens on screen. The oldest messages are simply dropped. The model does not tell you, and it does not know they are gone. What you notice is the assistant contradicting something you agreed twenty messages ago, or forgetting an instruction it followed perfectly earlier. That is almost always context loss rather than the model being unreliable.