Frequently asked questions

For GPT models, yes. Calculate Token runs OpenAI's own tokenizer in your browser, so a GPT count matches what OpenAI would bill for the same text. For Claude and Gemini it is an estimate: neither Anthropic nor Google publishes a tokenizer that can run in a browser, so those counts are derived from character length and every one of them is labelled "est" in the table. Estimates are usually close, but treat them as estimates.

Each provider uses its own tokenizer, so identical text splits into a different number of tokens on each one. Google's Gemini models tokenize differently from OpenAI's and Anthropic's, so a prompt that is 1,000 tokens on GPT may be somewhat higher or lower on Gemini. That is expected rather than an error, and it is one reason to compare cost per model rather than assuming one token count applies everywhere.

Counting never leaves your browser — GPT counts are computed locally with the real tokenizer, and Claude and Gemini counts are estimated locally from character length. One feature is different, and it is opt-in: Smart rewrite sends your prompt to Google's Gemini for a tighter version. Nothing else transmits what you typed unless you are signed in and choose to save a calculation, which stores the prompt text until you delete it.

The catalogue was last verified on 2026-08-31. It records when the figures were checked against every provider's own pricing page, not that today's page still matches them — providers change rates, run promotions and price long-context requests differently, sometimes without much notice. If that date is more than 30 days old, the site says so wherever the date is shown. Check the provider's own pricing page before committing to a number; every model page links to it.

Three things account for most of the gap, and none are visible in a block of text. System prompts are charged on every call, so a 600-token system prompt is billed whether or not the user's message is short. Conversation history compounds, because every prior turn is resent as input on the next one. And retries are billed — a failed parse or a client timeout pays for the request twice. Calculate Token prices the text you paste, which is a floor rather than a forecast.

On some models, sharply. The GPT-5.6 family bills the entire request at twice the input rate and one and a half times the output rate once input passes 272,000 tokens, and Gemini 3.1 Pro does the same above 200,000. It is a cliff rather than a gradient: a request just over the line costs nearly double one just under it. The comparison table applies the elevated rate automatically and marks it "long" when it does.

No. Counting, cost comparison, the model pages and the comparisons all work signed out, and signed out nothing you type is stored. An account only adds saving calculations, history, collections and exports.

Yes, and there is no metered tier. The site is funded by advertising. Token counting runs in your browser and costs nothing to serve; Smart rewrite calls a paid API, which is why that one feature is rate-limited.

A token is the unit a language model actually reads and the unit providers bill for. It is usually a word fragment rather than a word: common words are often one token, while rarer words, code and punctuation split into several. As a rough guide one token is about four characters of English, which is why a character count is a usable estimate but not a substitute for counting.

Prompt tokens are the text you send; completion tokens are what the model writes back. Providers bill them at different rates, and output is usually the dearer of the two — often three to six times the input rate. Calculate Token counts the text you paste, which is the prompt side, and prices the output side from the length you expect rather than from a real response, because that response does not exist yet. Switch the table between input and output pricing to see each half.

Shorten the prompt, cap the response length, pick the cheapest model that still does the job, and reuse a rewritten prompt rather than paying to rewrite it each time. Capping output is usually the bigger lever: output costs several times more per token than input, so trimming a 500-token answer saves more than trimming 500 tokens of prompt. The prompt optimizer does the first of these — Quick clean-up runs in your browser, and Smart rewrite sends your prompt to Google's Gemini.

Yes, with a free account. Saved calculations, collections and history persist between visits, and you can export everything as CSV, JSON, Markdown or a ZIP of all four. Exports can be imported again from the saved calculations page, which accepts the JSON and Markdown files this site produces — so a backup is something you can actually restore, not just download.