Token Calculator: Count Tokens and API Cost

A token calculator counts how many tokens your text uses in an AI model and turns that count into an API cost. Paste a prompt, a chat request or a file, pick a model from GPT, Claude, Gemini, Grok, DeepSeek or Mistral, and see your token count, context window use and the price per request and per month.

What Is a Token in AI, and What Is Tokenization?

A token is the small chunk of text an AI model reads and writes, and tokenization is the step that splits your text into those chunks. A token can be a whole short word, part of a long word, a number, a punctuation mark or a line break.

Models do not see letters or words. They see a list of token IDs, and every limit and price is based on that list. Your context window is measured in tokens, and API providers bill per million tokens.

For ordinary English, a widely used rule of thumb is that one token is about four characters, or about three quarters of a word. So 1,000 words come to roughly 1,300 tokens. Code, tables, URLs, emoji and languages such as Chinese, Japanese or Arabic usually split into more tokens per word, so the same idea can cost more.

Each model family has its own tokenizer, so the same text gives slightly different counts in different models. This is why a single rule of thumb is useful for planning, but a real count is better for budgeting.

How to Use the Token Calculator

You can get a token count and a cost estimate in a few steps, and you do not need an account to try it.

1. Type or paste your text into the box labeled Your text, prompt or chat request (JSON). You can also use Paste, press Sample to load an example, or upload a file. Supported files include text, Markdown, CSV, JSON, code and .docx, up to 20 MB.
2. If you pasted a messages array from OpenAI, Claude or Gemini, keep Count it as a chat request checked. The count then includes the formatting tokens each message adds. Uncheck it to count the raw text only.
3. Choose a model in the AI model list. Models are grouped by provider, and Include older models adds models that are still on sale but no longer current.
4. Check the Token count and the Context window used bar to see how much of the model's limit your text fills.
5. In the cost section, leave Input tokens per request on Use the tokens of my text, or type your own number. Set Output tokens per request manually, choose Same as input (rewrites, translations), or use the typical reply lengths shown.
6. Enter the Number of requests and switch the period between per month and per day. Add a Cached input (prompt caching), % value if part of your prompt repeats, and tick Batch API if you plan to use batch pricing.
7. Read the comparison table. It lists each model with its tokens, cost per request, cost for your chosen period and context size, and you can sort it or limit it with Providers to compare.

Worked Example: Estimating a Monthly Bill

Say you run a support assistant that sends the same 1,500-word help article plus a customer question with every request. That is about 2,000 input tokens per request, and you expect replies of about 500 tokens. You expect 10,000 requests a month.

For the math, assume a model priced at $2 per 1M input tokens and $8 per 1M output tokens. These prices are only an example, so check the live prices in the table for the model you choose.

• Input: 2,000 tokens × 10,000 requests = 20,000,000 tokens, which is 20 × $2 = $40.
• Output: 500 tokens × 10,000 requests = 5,000,000 tokens, which is 5 × $8 = $40.
• Total: about $80 per month, or $0.008 per request.

Now suppose the 1,500-word article is the same in every call. If you set Cached input (prompt caching), % to a high share, the cached part is billed at the lower cached-input rate shown in the price table, and the monthly figure drops. Output tokens often cost several times more than input tokens, so a long reply can cost more than a long prompt.

How Do You Read and Act on the Results?

Each result answers one planning question: does the text fit, what does it cost, and which model is the best value for the job.

Token count and the exact or estimate badge

The big number is the token count for the selected model. A badge tells you whether it is an exact count or an estimate. Use exact counts for tight limits, and treat estimates as a planning figure with a small margin of error.

Context window used

The bar shows your tokens as a share of the model's context window. Below about 70% you have comfortable room. Once the bar turns amber, you may leave too little space for the reply. A red bar means the text is over the limit and must be shortened or split. Remember that the model's answer also uses the same window, so leave space for output.

Cost per request and per month

Cost per request is useful for comparing prompts. Cost for the period is useful for budgets. If the monthly number is higher than you expected, try these fixes in order:

• Trim repeated instructions and long examples from the prompt.
• Move stable text to the front so caching applies, then raise the cached share.
• Cap the reply length so you do not pay for output you do not use.
• Use batch pricing for work that does not need an instant answer.
• Compare a smaller model in the table and test whether its quality is enough.

Words, characters and the words-per-token ratio

The tool also shows characters and words, plus a quick answer to How many words is a token? for your own text. A ratio far from the usual 0.75 words per token is a hint that your text has a lot of code, numbers or non-English content.

How Can You See How Your Text Splits Into Tokens?

The section See how your text splits into tokens shows your text with each token highlighted in alternating colors. Choose a Tokenizer: o200k_base is used by GPT-5, GPT-4.1, GPT-4o and the o-series, and cl100k_base is used by GPT-4, GPT-4 Turbo and GPT-3.5.

Turn on Show spaces and line breaks to see that a leading space is often part of a token. Turn on Show token IDs to see the number the model actually receives. This view explains surprises, such as why an unusual product name takes five tokens while a common word takes one.

Who Is a Token Calculator For?

It helps anyone who writes prompts or pays for AI usage, and it saves guessing before you send a request.

• Developers check that a prompt and its documents fit inside a context window before an API call fails.
• Product teams forecast a monthly bill for a chatbot, summarizer or translation feature before launch.
• Content and SEO teams estimate the cost of rewriting or summarizing hundreds of pages in bulk.
• Prompt engineers measure how much shorter a rewritten prompt really is.
• Analysts compare providers side by side for the same workload using the same inputs.

Best Practices, Common Mistakes and Troubleshooting

Most wrong estimates come from a few predictable input mistakes, and each has a simple fix.

• Forgetting the output. Cost depends on input and output. If you only count the prompt, your bill will look too low. Set a realistic output size.
• Ignoring chat overhead. A conversation adds formatting tokens to every message. Paste the full messages array and keep Count it as a chat request checked.
• Counting only one turn. In a multi-turn chat, earlier messages are sent again with each new request. Enter the full history size as your input tokens.
• Using the wrong tokenizer. If you compare GPT counts, make sure the tokenizer matches the model generation.
• Pasting JSON that is not a messages array. If your JSON is data and not a chat, uncheck the chat option so it is counted as plain text.
• Hidden text in files. A .docx or HTML file can contain more text than you see on screen. Check the character count when a number looks too high.
• Old prices. Providers change prices. Look at the Prices checked note and confirm against the provider before signing off a budget.

How Accurate Is This Token Calculator?

OpenAI counts use the real tokenizers and are exact for the models that share them. For Claude, Gemini, Grok, DeepSeek and Mistral, a count may be an estimate scaled from the GPT count, and the badge says so. Estimates are good for budgeting, but they can differ from the provider's own figure by a few percent.

For very large texts, above 200,000 characters, the character and word counts use a faster method and can differ slightly from a word processor. Prices and context sizes are a snapshot, so they can lag behind a provider's announcement. Finally, a token count is not a quality measure, and a cheaper model is only a better choice if it still does the job.

Paste your own prompt or load the sample in the calculator above, pick a model, and you will see your token count and cost in seconds. Then adjust the output size, request volume and caching to see what your real bill could look like.