How Many Tokens Is Your Prompt? How to Count Tokens Before You Send

Every AI request you send is measured in tokens, and those tokens decide both your API bill and whether your text even fits in the model’s memory. The good news is that you can count them in seconds before sending. Paste your text into our free AI Token Counter and you will get an estimated token count, character count, word count, and a clear view of how much of each model’s context window your text uses. This guide explains the counting rules behind it.

What exactly is a token?

A token is a small chunk of text that an AI model reads or writes, roughly three quarters of an English word. Models split text into tokens using a dictionary of common pieces: short words are usually one token, long or unusual words split into several, and punctuation, spaces, numbers and code all count too.

Because tokenisation is mechanical rather than linguistic, token counts feel unintuitive at first. A 750-word article is about 1,000 tokens. A single emoji can be two tokens. A long URL can be dozens. You do not need to memorise any of this; you just need a quick way to measure, which is what a token counter is for.

How many tokens is a typical prompt?

A short, single question is usually 15 to 30 tokens. A well-built working prompt with instructions and an example runs 200 to 800 tokens. A prompt that pastes in a document, a web page or a long brief easily reaches 2,000 to 10,000 tokens, and chat history can push a request far higher.

The spread matters because cost and limits both scale with tokens. The same question can be a 30-token request typed fresh, or a 30,000-token request by the time a chatbot resends fifty previous messages along with it. When people say an AI tool got expensive or started forgetting things, oversized prompts are usually the hidden reason.

How do you count tokens before sending?

The quick way is a token counter: paste the text and read the estimate. The counting rules behind it are simple. English text averages about 4 characters per token and about 1.33 tokens per word, so dividing characters by 4, or multiplying words by 1.33, both give a fair estimate; averaging the two is closer still.

For an exact figure, use your provider’s own tokeniser, because GPT, Claude and Gemini each split text slightly differently and counts can vary by a few percent between families. Estimates are fine for budgeting, trimming and checking limits. Only reach for the exact tokeniser when you are reconciling an invoice or squeezing under a hard limit.

What is a context window, and how much of it are you using?

A context window is the maximum amount of text a model can consider at once, measured in tokens, and it must hold your prompt, any documents, the conversation history and the model’s answer together. Common sizes are 128K tokens for GPT-4o class models, 200K for Claude class models, 400K for GPT-5 class models and 1M for Gemini class models.

Our AI Token Counter shows your text as a percentage of each of these windows, with the room left over. Keep a healthy margin: if your prompt fills 95 percent of the window, there is almost nothing left for the answer, and the model may cut off mid-reply or drop your earliest text without warning.

Why do token counts differ between models?

Each provider trains its own tokeniser with its own dictionary of text pieces, so the same paragraph produces slightly different counts on GPT, Claude and Gemini, usually within a few percent for English prose. The gap widens for code, maths, URLs and languages that use non-Latin scripts, which split into many more tokens per word.

Practically, this means a count taken on one model is a guide for the others, not gospel. If your prompt sits comfortably under every window size and your budget has slack, the small differences do not matter. If you are working right at a limit, count again with the specific model’s tokeniser.

How can you make a prompt use fewer tokens?

Cut instruction filler, delete duplicated examples, and say each rule once. Summarise long background documents instead of pasting them whole, drop finished conversation turns, and remove formatting noise such as repeated blank lines and pasted web page clutter. Most real prompts shrink by a third without losing anything the model actually uses.

Then measure again. Shorter prompts cost less on every request, respond faster, and leave more of the context window free for the answer and for new information. If you pay per token, price your before-and-after counts in the AI Token Cost Calculator to see what the tidy-up saves per month.

Frequently asked questions

Is 1,000 tokens a lot?

It is about 750 words, roughly a long email or a short blog post. As a prompt it is modest; as a per-request cost repeated thousands of times a day, it is a real budget line worth measuring.

Do spaces and punctuation count as tokens?

Yes, in effect. Spaces and punctuation are bundled into the token pieces or counted as their own tokens, which is why the character-based rule of 4 characters per token works so well as an estimate.

Why does my chatbot forget the start of long chats?

Because the full history is resent with every message until it fills the context window, at which point the oldest turns are dropped. Summarising the conversation, or starting a fresh chat with a summary, fixes it.

Can I count tokens for a whole book?

Yes. A typical novel is 100,000 to 150,000 tokens, which fits whole in the larger context windows but not in a 128K window once you add instructions and the answer. The counter will show you exactly where your text lands against each window.

Token counts from any estimator are approximate, because each model family tokenises text slightly differently; use your provider’s tokeniser for exact billing figures.