Count tokens, estimate API usage, optimize prompts, and discover ways to reduce unnecessary token consumption.
A short, practical guide to what the numbers above mean and how to use them when you build with large language models.
Models don't read letters or words, they read tokens: pieces of text such as a whole word, part of a word, a number, punctuation or even a fragment of an emoji. In English a token is roughly ¾ of a word (about 4 characters), but that ratio changes a lot across languages, code and formatting.
Each model family has its own tokenizer with its own vocabulary, so the same text produces a different count. Neither number is “wrong”. OpenAI models are counted exactly here; for other providers, which don't publish an offline tokenizer, the tool shows a clearly marked estimate.
APIs bill per million tokens, with separate prices for input and output. Cost = input tokens × input price + output tokens × output price. Some models charge more for very long prompts, and some prices are introductory, so check the provider's page before budgeting.
The context window is the most text a model can consider at once: your prompt, any attached data and the reply it writes all share it. If your input plus the expected output is larger than the window, the request fails or gets truncated. The meter above shows how much of the window you are using.
Input tokens are what you send. Output tokens are what the model generates, which we can't know in advance, so enter the reply length you expect. Output is usually priced several times higher than input, which is why long answers dominate many bills.
The count above covers the text you entered. A real request also carries system instructions, conversation history, attached files, tool definitions and the reply, so an IDE or API dashboard can report more. Neither number is wrong: they measure different things.
An AI model rewrites your prompt to remove repetition and filler while keeping requirements, constraints and examples. A second pass checks that nothing important was lost, and both versions are recounted locally with the selected model's tokenizer, so the savings are measured rather than promised. You always review the result before it replaces your text.
Related tools: LLM Cost Calculator, Context Window Checker, JSON Formatter and PDF to Text (to extract text from a PDF before counting).
A prompt that looks short can still overflow a context window or surprise you on the bill, because both are measured in tokens, not words. Paste text, JSON, Markdown or code, or load a text file, and the count updates as you type for the model you picked. Each result is labeled Exact or Estimated, so you know whether it comes from the model's own tokenizer or from a stand-in. Context use and an input cost estimate sit right beside it.
Example input
Summarize this JSON API response in two sentences for a non-technical stakeholder.