Guide · 8 min read · free tool
Token Counts: A Rough Length, Not the Bill the API Will Send
Sep 1, 2026 · Zelpura Editorial · All guides
Models do not read 'words' the way a teacher does. They read tokens, which are chunks of text, and they bill and limit by those chunks. A token is often around four characters in English and a worse deal in some other languages and in code. This page estimates a count from the text you paste, locally, so you can see whether a prompt is a paragraph or a novella before you hit an API. It is an estimate. The provider's tokenizer is the invoice. If the two disagree, the invoice wins, and you should not be surprised that a sketch lost an argument with a meter.
The inputs are the job, not a formality. Paste the text you are about to send, or a representative sample if the real text is huge. The page estimates GPT at about 4 characters per token, Claude at about 3.5, and Gemini at about 4.2, then shows a rough cost sketch beside the counts. The count is for what you pasted, not for the hidden system prompt, the tools, the conversation history, or the answer you have not received yet. People compare a token count of their question to a context window and forget that the window also has to hold the reply. Leave room. A 100,000 token window is not a 100,000 token prompt. It is a budget for prompt, history, and completion. This page does not know your history. Subtract it yourself.
A 2,000 word essay in English often lands in the neighborhood of 2,500 to 3,000 tokens, and a table or a pile of code can land somewhere else entirely. If your limit is 8,000 and your pasted prompt estimates at 6,000, you do not have 2,000 tokens of comfort. You have 2,000 minus the system prompt minus the answer. Ask for a short answer or cut the prompt. Cutting is usually possible. Most prompts contain a second copy of the instructions and a paragraph of apology. Delete those before you buy a bigger window. The counter is how you notice the paragraph. Your eye had gotten used to it.
The miss is treating the estimate as the bill. Tokenizers differ. A count that is 15 percent off is still useful for planning and useless for a dispute with a provider. The second miss is counting only the user message. The third miss is pasting secrets into any counter, including this one. This one runs locally, which is better than a random website, and your clipboard still has the secret. Do not paste keys, medical notes, or client text into a tool you do not need to debug. Count a dummy of the same length if the content is sensitive. The length is the input. The content is none of the counter's business.
Paste a non-sensitive sample. Read the estimate. Compare it to the context you actually have, after system prompt and expected reply. If you are over, cut. If you are choosing between models, use the count as a relative size, then check the provider's own tokenizer before you sign a contract that depends on the cost. For repeated prompts, count the template once and the variable part as a range. A template that is already 2,000 tokens is a design problem. Shorten the template. You will pay it on every call, which is the definition of a cost worth editing.
This is a rough local estimate, not a provider's tokenizer and not a bill. It is not advice to send any particular text to any particular model. Context limits and prices change. A count does not tell you whether the model will follow the instructions, refuse the request, or leak the prompt into a log on the provider's side. Read the provider's data terms before you paste anything you would not put in an email to them. The fact that you counted it here does not change their retention. Counting and sending are different steps. Do not collapse them.
If you do this often, write down the estimate-versus-invoice gap for your kind of text. A personal fudge factor beats rediscovering that your code counts heavy. Apply the fudge next time. When a teammate says a prompt is 'just a few thousand words,' ask for the token estimate and the reserved reply. The argument gets shorter. Shorter arguments are the hidden benefit of a counter. The visible benefit is not blowing a context limit on a demo day.
Read the result back before you act. Say the estimate, then say what you are reserving for the answer. If the second number is zero, you have planned a prompt that leaves the model no room to speak. Leave room.
Images, audio, and tool outputs are not free just because this box counted text. A multimodal request has other meters. Do not use a text token count to approve a workflow that sends files. Price the files separately, with the provider's units, and keep customer files out of casual tests. A test with real customer data is not a test. It is a disclosure. Use synthetic text of the same length. The counter cannot tell the difference, which is exactly why you have to.
Related: Sub vs API · Print cost · Newsletter CPM · All 93 tools