Tokens ↔ Words ↔ Characters Converter
Convert between tokens, words and characters for LLM prompts. Type any one and get the other two instantly. Free, runs in your browser.
Tokens, Words and Characters Converter
Type any one value — the other two update instantly. Choose the model and the writing system first: characters per token differ several-fold between English and Chinese, Japanese, Korean or Thai. These are estimates; for an exact GPT count, use the Token Counter.
How to Use the Converter
Type into any box
Pick the model and the writing system, then enter a number of tokens, words, or characters. You only ever fill one — the tool works out the other two. Commas between thousands are fine.
Read the conversion
The other two boxes update instantly. For English on the GPT setting that is 4 characters per token and 5.2 characters per word, so 500 words comes to 650 tokens. Chinese, Japanese, Korean and Thai are converted by characters only: they are not written with spaces, so there is no word count to convert.
Use it to size a prompt
Heading into a model with a token budget or a context window? Convert your word count to tokens to see roughly how much room you'll use.
Switch to exact when it matters
These are estimates, set to err towards more tokens rather than fewer. For billing or a tight context window, count the real text with the Token Counter, which runs the actual GPT tokenizer.
Tokens, Words and Characters: A Quick Mental Model
Why a rough conversion is so useful
You usually think in words — "a 500-word email", "a 2,000-word article" — but language models think in tokens, and that's the unit of both their context limit and their price. A fast conversion lets you answer everyday questions without pasting anything into a tokenizer: will this article fit in the prompt? Roughly how much will this batch cost? The vendors publish their own rules of thumb. Google says that for Gemini "a token is equivalent to about 4 characters" and "100 tokens is equal to about 60-80 English words". Anthropic says a Claude token "represents approximately 3.5 English characters" on models before Claude 4.7, and that the newer tokenizer in 4.7 and later "produces approximately 30 percent more tokens for the same text".
We measured OpenAI's o200k tokenizer (GPT-4o, GPT-4.1, GPT-5) on two English texts: the first chapter of Pride and Prejudice came to 0.81 words per token, and this site's About page to 0.79 — one ran 5.2 characters per word, the other 6.1. This converter uses 5.2 characters per word and the Token Counter's GPT rate of 4 characters per token, so 2,000 words reads as 2,600 tokens: a little more than either text actually used, which is the safe side for a budget.
"Think in words, budget in tokens. A good estimate bridges the two in your head — an exact count settles the invoice."
Where the estimate drifts
Tokenizers split common words into a single token and break rare or long words into pieces, and they count spaces and punctuation, so code, URLs, numbers and identifiers usually take more tokens than prose of the same length. The biggest difference is the writing system: on the GPT setting the converter counts about 4 characters per token for English but about 1 for Chinese, which is why Chinese, Japanese, Korean and Thai each have their own rate here. Claude is the exception — Anthropic publishes a figure for English only, so for other scripts this tool shows no Claude estimate rather than guess. When a few percent matters for your bill or your context window, measure the real text.
Where the estimate drifts
The 4-characters-per-token figure holds up well for ordinary English writing. It starts to drift in predictable directions: dense code and JSON push the token count up relative to characters because of symbols and indentation; long technical terms and identifiers split into multiple tokens; and languages with non-Latin scripts can use dramatically more tokens per character. If your text is mostly one of those, nudge your mental ratio accordingly — or just measure it. The point of this tool is to get you 90% of the way there in a second, so you only spend the effort of an exact count when the last few percent actually changes a decision.
Code and JSON burn more tokens per character than prose
OpenAI's o200k tokenizer measured 0.79–0.81 words per token on two English texts — a novel chapter and a web page.
Google says a Gemini token is about 4 characters, and 100 tokens about 60–80 English words.
Anthropic puts a Claude token at about 3.5 English characters before Claude 4.7 — and says 4.7 and later produce about 30% more tokens for the same text.
On the GPT setting, Chinese runs about 1 character per token against 4 for English — so a character count alone hides a fourfold difference.
Code and JSON usually use more tokens per character than prose, because of symbols, indentation and long identifiers.
Spaces and punctuation are tokens too, which is why a character count that includes them is the steadier basis.
Context windows and API prices are both measured in tokens, not words.
These conversions are estimates — the only exact count comes from running the model's own tokenizer.
Characters per word is not a constant: the two English texts we measured ran 5.2 and 6.1. The converter uses the lower, so a word count never converts to too few tokens.
This tool runs entirely in your browser — nothing is sent anywhere.
Frequently Asked Questions
- On this converter's GPT setting, 769 words of English: 1,000 tokens × 4 characters per token ÷ 5.2 characters per word. On the two English texts we measured with OpenAI's tokenizer, 1,000 tokens held 785–814 words, so 769 errs towards fewer words than will really fit. Gemini gives 788; Claude before 4.7 gives 673.
- On the GPT setting, 650 tokens: 500 × 5.2 characters per word ÷ 4 characters per token. On Claude 4.7 and later the same 500 words reads as 966. Dense code, long technical terms or non-English text push the real number higher.
- No — it's an average-based estimate. The only way to get an exact count is to run the model's tokenizer on the actual text. For OpenAI models you can do that with our Token Counter, which uses the real tiktoken encodings.
- A tokenizer gives common character sequences a single token and breaks rarer ones into pieces. Symbols, indentation and long identifiers are rarer than ordinary words, so they split more; and each writing system has its own rate — on the GPT setting about 4 characters per token for English, about 1 for Chinese and about 1.4 for Japanese kana or Korean.
- Yes, by enough to matter for a budget, which is why the converter asks for the model. For English: 4 characters per token on the GPT setting, 4.1 for Gemini, 3.5 for Claude before 4.7 (Anthropic's figure), and about 2.7 for Claude 4.7 and later, which Anthropic says produces about 30% more tokens for the same text. For a precise per-model figure you need that model's own tokenizer.
- Because you often only know a word count — "a 2,000-word article" — and want a quick sense of whether it fits a context window or how much it might cost, without pasting the text anywhere. This converter gives that instant ballpark; the Token Counter gives the exact figure when you have the text.
- Yes. Type into tokens, words or characters and the other two update. It's fully bidirectional, so you can start from whichever number you already have.
- No. It's just arithmetic running in your browser. The numbers you type are never sent to any server or third party, and nothing is stored.
- This tool only converts counts. To turn token counts into a dollar figure, use our LLM Cost Calculator, which applies each model's input and output rates to your token numbers.
- Characters per token come from the same rates our Token Counter uses: for GPT, OpenAI's o200k tokenizer, set never to count fewer tokens than it measured; for Gemini, a fit to Google's own countTokens endpoint (October 2026), close to Google's published "about 4 characters". For Claude we use Anthropic's published English figure only. Characters per word, 5.2, we measured on two English texts. All of it is a planning estimate; use the tokenizer of the model you are billed on when it matters.
Method & sources
How it computes
Estimate: tokens = characters ÷ characters-per-token for the chosen model and writing system (GPT: OpenAI o200k-based rates; Gemini: fitted to Google countTokens; Claude: Anthropic's 3.5 English characters, +30% tokens on 4.7+, English only). Words = characters ÷ 5.2, measured on English with o200k; no word count for scripts without spaces. Tokens round up, words and characters down (TokenEstimateKernel.convertCount).
What this tool implements
- Characters per token by model AND writing system (Han, kana, Hangul, Thai, Latin), because the rate differs several-fold between scripts: the GPT rates are the Token Counter's o200k-based rates (never below o200k), the Gemini rates its fit to Google's countTokens endpoint (October 2026).
- Claude: Anthropic's published figure only — about 3.5 English characters per token before Claude 4.7, and about 30% more tokens for the same text on 4.7 and later. Anthropic publishes no figure for other scripts, so the page shows no Claude estimate for them rather than guess.
- The Claude rates are the Token Counter's Claude rates, read from the same kernel, so both pages give the same token count for the same English text on both Claude families.
- Characters per word measured, not quoted: OpenAI o200k on two English texts, 2026-10-19 — Pride and Prejudice ch. 1 (5.21 chars/word, 0.814 words/token) and the site's About page (6.07, 0.785). The lower, 5.2, is used so a word count never converts to fewer tokens than o200k measured on either.
- Rounding is on the safe side for a budget: tokens round up from words or characters; words and characters round down from tokens. Results are labelled estimates; exact counts need the billed model's tokenizer.
- Corrected 2026-10-19: one English ratio (0.75 words, 4 characters per token) was applied to every language, including Chinese, Japanese, Korean and Thai; it was cited to an OpenAI Help Center page that refuses a raw fetch (403) and is no longer cited; "500 words is 650–700 tokens" and "1,000 tokens is about 750 words" are now the converter's own 650 and 769; "tokenizers are trained mostly on English" was unsourced.
Sources
- Anthropic. Glossary — Tokens: a token "represents approximately 3.5 English characters" on models before Claude 4.7; the newer tokenizer "produces approximately 30 percent more tokens for the same text": https://docs.anthropic.com/en/docs/about-claude/glossary
- Google AI for Developers. Understand and count tokens — "a token is equivalent to about 4 characters. 100 tokens is equal to about 60-80 English words": https://ai.google.dev/gemini-api/docs/tokens
- OpenAI. tiktoken (o200k_base, cl100k_base), the tokenizer used for the measurement, via the vendored js-tiktoken: https://github.com/openai/tiktoken
- Austen J. Pride and Prejudice (1817), Chapter 1 — measurement text: https://en.wikisource.org/wiki/Pride_and_Prejudice_%281817%29/Cha…
Abridged — the full review record for this tool runs longer than the list above.
Pick up where you left off
Stored only in this browser — never sent to our servers.