Token Counter
Token Counter
Paste text to count tokens with each model's real tokenizer, plus character count, word count, and estimated cost.
TL;DR
This counts tokens using each model's real tokenizer — exact for OpenAI models (o200k_base), approximate for others. As a rule of thumb, English prose runs about 0.75 words per token (≈ 1.3 tokens per word, ≈ 4 characters per token); code is denser. Paste text above for an instant count; Claude counts are approximate unless you use Anthropic's count-tokens API.
How many tokens is X?
Updated Jul 5, 2026A token is the unit language models read and bill by — roughly a common word or word-piece, not a character or a whole word. As a rule of thumb, English prose runs about 0.75 words per token (≈ 1.3 tokens per word, ≈ 4 characters per token). Code is denser: punctuation, indentation, and long identifiers push the token count higher per character.
| Input | Approx. tokens |
|---|---|
| 1 word (English prose) | ~1.3 |
| 1 sentence (~15 words) | ~20 |
| 1 paragraph (~100 words) | ~130 |
| 1 page of prose (~500 words) | ~650 |
| 1,000 words | ~1,300 |
| A ~200-line code file | ~1,500–2,500 |
| A short novel (~50,000 words) | ~65,000 |
Counts use the GPT-4o / GPT-5 tokenizer (o200k_base). Other model families use different tokenizers, so exact counts vary — see the counter above for a per-model estimate.
How it works
OpenAI models are tokenized exactly in your browser with gpt-tokenizer (the o200k_base encoding used by GPT-4o, GPT-4.1, GPT-5, and the o-series). Counting is debounced and runs fully client-side — your text is never uploaded.
For models with non-public or different tokenizers (Claude, Gemini, Llama, Mistral, DeepSeek, Grok), the count is a labelled approximation derived from the GPT-4o count with a per-family multiplier (Claude ≈ 1.18×, others ≈ 1.05–1.10×). The estimated cost multiplies the token count by the selected model's input rate from the cost calculator data layer.
FAQ
- How many tokens is a word?
- For English prose, about 1.3 tokens per word on average (roughly 0.75 words per token, or ~4 characters per token). So 100 words is ~130 tokens and a 500-word page is ~650 tokens. Code and non-English text are denser and produce more tokens per word.
- Is this token count exact?
- For OpenAI models (GPT-4o, GPT-4.1, GPT-5, o-series) the count is exact — it runs the real o200k_base tokenizer in your browser. For Claude, Gemini, Llama and others the count is a labelled approximation, because those models use different tokenizers. For an exact Claude count, use Anthropic's count-tokens API.
- Why does Claude report a different token count than GPT?
- Each model family uses its own tokenizer with a different vocabulary, so the same text splits into different numbers of tokens. Claude's tokenizer typically yields ~15–20% more tokens than OpenAI's tiktoken for the same English text, which is why this tool scales the GPT count up for Claude estimates.
- Does the token count include the system prompt and formatting?
- This tool counts exactly the text you paste. Real API requests also spend tokens on the system prompt, message-role formatting, tool/function definitions, and (on reasoning models) hidden thinking tokens — so a full request is larger than the visible user text alone.
- Is my text sent anywhere?
- No. Counting runs entirely in your browser using a local copy of the tokenizer. Nothing you paste is uploaded to a server.